[Features] Support PV_RCNN modules (#1957)

* add pvrcnn module code

* add voxelsa

* fix

* fix comments

* fix comments

* fix comments

* add stack sa

* fix

* fix comments

* fix comments

* fix

* add ut

* fix comments

configs/pvrcnn/pvrcnn_8xb2-80e_kitti-3d-3class.py

+_base_ = [
+    '../_base_/datasets/kitti-3d-3class.py',
+    '../_base_/schedules/cyclic-40e.py', '../_base_/default_runtime.py'
+]
+
+voxel_size = [0.05, 0.05, 0.1]
+point_cloud_range = [0, -40, -3, 70.4, 40, 1]
+
+data_root = 'data/kitti/'
+class_names = ['Pedestrian', 'Cyclist', 'Car']
+metainfo = dict(CLASSES=class_names)
+db_sampler = dict(
+    data_root=data_root,
+    info_path=data_root + 'kitti_dbinfos_train.pkl',
+    rate=1.0,
+    prepare=dict(
+        filter_by_difficulty=[-1],
+        filter_by_min_points=dict(Car=5, Pedestrian=5, Cyclist=5)),
+    classes=class_names,
+    sample_groups=dict(Car=15, Pedestrian=10, Cyclist=10),
+    points_loader=dict(
+        type='LoadPointsFromFile', coord_type='LIDAR', load_dim=4, use_dim=4))
+
+train_pipeline = [
+    dict(type='LoadPointsFromFile', coord_type='LIDAR', load_dim=4, use_dim=4),
+    dict(type='LoadAnnotations3D', with_bbox_3d=True, with_label_3d=True),
+    dict(type='ObjectSample', db_sampler=db_sampler, use_ground_plane=True),
+    dict(type='RandomFlip3D', flip_ratio_bev_horizontal=0.5),
+    dict(
+        type='GlobalRotScaleTrans',
+        rot_range=[-0.78539816, 0.78539816],
+        scale_ratio_range=[0.95, 1.05]),
+    dict(type='PointsRangeFilter', point_cloud_range=point_cloud_range),
+    dict(type='ObjectRangeFilter', point_cloud_range=point_cloud_range),
+    dict(type='PointShuffle'),
+    dict(
+        type='Pack3DDetInputs',
+        keys=['points', 'gt_bboxes_3d', 'gt_labels_3d'])
+]
+test_pipeline = [
+    dict(type='LoadPointsFromFile', coord_type='LIDAR', load_dim=4, use_dim=4),
+    dict(
+        type='MultiScaleFlipAug3D',
+        img_scale=(1333, 800),
+        pts_scale_ratio=1,
+        flip=False,
+        transforms=[
+            dict(
+                type='GlobalRotScaleTrans',
+                rot_range=[0, 0],
+                scale_ratio_range=[1., 1.],
+                translation_std=[0, 0, 0]),
+            dict(type='RandomFlip3D'),
+            dict(
+                type='PointsRangeFilter', point_cloud_range=point_cloud_range)
+        ]),
+    dict(type='Pack3DDetInputs', keys=['points'])
+]
+
+model = dict(
+    type='PointVoxelRCNN',
+    data_preprocessor=dict(
+        type='Det3DDataPreprocessor',
+        voxel=True,
+        voxel_layer=dict(
+            max_num_points=5,  # max_points_per_voxel
+            point_cloud_range=point_cloud_range,
+            voxel_size=voxel_size,
+            max_voxels=(16000, 40000))),
+    voxel_encoder=dict(type='HardSimpleVFE'),
+    middle_encoder=dict(
+        type='SparseEncoder',
+        in_channels=4,
+        sparse_shape=[41, 1600, 1408],
+        order=('conv', 'norm', 'act'),
+        encoder_paddings=((0, 0, 0), ((1, 1, 1), 0, 0), ((1, 1, 1), 0, 0),
+                          ((0, 1, 1), 0, 0)),
+        return_middle_feats=True),
+    points_encoder=dict(
+        type='VoxelSetAbstraction',
+        num_keypoints=2048,
+        fused_out_channel=128,
+        voxel_size=voxel_size,
+        point_cloud_range=point_cloud_range,
+        voxel_sa_cfgs_list=[
+            dict(
+                type='StackedSAModuleMSG',
+                in_channels=16,
+                scale_factor=1,
+                radius=(0.4, 0.8),
+                sample_nums=(16, 16),
+                mlp_channels=((16, 16), (16, 16)),
+                use_xyz=True),
+            dict(
+                type='StackedSAModuleMSG',
+                in_channels=32,
+                scale_factor=2,
+                radius=(0.8, 1.2),
+                sample_nums=(16, 32),
+                mlp_channels=((32, 32), (32, 32)),
+                use_xyz=True),
+            dict(
+                type='StackedSAModuleMSG',
+                in_channels=64,
+                scale_factor=4,
+                radius=(1.2, 2.4),
+                sample_nums=(16, 32),
+                mlp_channels=((64, 64), (64, 64)),
+                use_xyz=True),
+            dict(
+                type='StackedSAModuleMSG',
+                in_channels=64,
+                scale_factor=8,
+                radius=(2.4, 4.8),
+                sample_nums=(16, 32),
+                mlp_channels=((64, 64), (64, 64)),
+                use_xyz=True)
+        ],
+        rawpoints_sa_cfgs=dict(
+            type='StackedSAModuleMSG',
+            in_channels=1,
+            radius=(0.4, 0.8),
+            sample_nums=(16, 16),
+            mlp_channels=((16, 16), (16, 16)),
+            use_xyz=True),
+        bev_feat_channel=256,
+        bev_scale_factor=8),
+    backbone=dict(
+        type='SECOND',
+        in_channels=256,
+        layer_nums=[5, 5],
+        layer_strides=[1, 2],
+        out_channels=[128, 256]),
+    neck=dict(
+        type='SECONDFPN',
+        in_channels=[128, 256],
+        upsample_strides=[1, 2],
+        out_channels=[256, 256]),
+    rpn_head=dict(
+        type='PartA2RPNHead',
+        num_classes=3,
+        in_channels=512,
+        feat_channels=512,
+        use_direction_classifier=True,
+        dir_offset=0.78539,
+        anchor_generator=dict(
+            type='Anchor3DRangeGenerator',
+            ranges=[[0, -40.0, -0.6, 70.4, 40.0, -0.6],
+                    [0, -40.0, -0.6, 70.4, 40.0, -0.6],
+                    [0, -40.0, -1.78, 70.4, 40.0, -1.78]],
+            sizes=[[0.8, 0.6, 1.73], [1.76, 0.6, 1.73], [3.9, 1.6, 1.56]],
+            rotations=[0, 1.57],
+            reshape_out=False),
+        diff_rad_by_sin=True,
+        assigner_per_size=True,
+        assign_per_class=True,
+        bbox_coder=dict(type='DeltaXYZWLHRBBoxCoder'),
+        loss_cls=dict(
+            type='mmdet.FocalLoss',
+            use_sigmoid=True,
+            gamma=2.0,
+            alpha=0.25,
+            loss_weight=1.0),
+        loss_bbox=dict(
+            type='mmdet.SmoothL1Loss', beta=1.0 / 9.0, loss_weight=2.0),
+        loss_dir=dict(
+            type='mmdet.CrossEntropyLoss', use_sigmoid=False,
+            loss_weight=0.2)),
+    roi_head=dict(
+        type='PVRCNNRoiHead',
+        num_classes=3,
+        semantic_head=dict(
+            type='ForegroundSegmentationHead',
+            in_channels=640,
+            extra_width=0.1,
+            loss_seg=dict(
+                type='mmdet.FocalLoss',
+                use_sigmoid=True,
+                reduction='sum',
+                gamma=2.0,
+                alpha=0.25,
+                activated=True,
+                loss_weight=1.0)),
+        bbox_roi_extractor=dict(
+            type='Batch3DRoIGridExtractor',
+            grid_size=6,
+            roi_layer=dict(
+                type='StackedSAModuleMSG',
+                in_channels=128,
+                radius=(0.8, 1.6),
+                sample_nums=(16, 16),
+                mlp_channels=((64, 64), (64, 64)),
+                use_xyz=True,
+                pool_mod='max'),
+        ),
+        bbox_head=dict(
+            type='PVRCNNBBoxHead',
+            in_channels=128,
+            grid_size=6,
+            num_classes=3,
+            class_agnostic=True,
+            shared_fc_channels=(256, 256),
+            reg_channels=(256, 256),
+            cls_channels=(256, 256),
+            dropout_ratio=0.3,
+            with_corner_loss=True,
+            bbox_coder=dict(type='DeltaXYZWLHRBBoxCoder'),
+            loss_bbox=dict(
+                type='mmdet.SmoothL1Loss',
+                beta=1.0 / 9.0,
+                reduction='sum',
+                loss_weight=1.0),
+            loss_cls=dict(
+                type='mmdet.CrossEntropyLoss',
+                use_sigmoid=True,
+                reduction='sum',
+                loss_weight=1.0))),
+    # model training and testing settings
+    train_cfg=dict(
+        rpn=dict(
+            assigner=[
+                dict(  # for Pedestrian
+                    type='Max3DIoUAssigner',
+                    iou_calculator=dict(type='BboxOverlapsNearest3D'),
+                    pos_iou_thr=0.5,
+                    neg_iou_thr=0.35,
+                    min_pos_iou=0.35,
+                    ignore_iof_thr=-1),
+                dict(  # for Cyclist
+                    type='Max3DIoUAssigner',
+                    iou_calculator=dict(type='BboxOverlapsNearest3D'),
+                    pos_iou_thr=0.5,
+                    neg_iou_thr=0.35,
+                    min_pos_iou=0.35,
+                    ignore_iof_thr=-1),
+                dict(  # for Car
+                    type='Max3DIoUAssigner',
+                    iou_calculator=dict(type='BboxOverlapsNearest3D'),
+                    pos_iou_thr=0.6,
+                    neg_iou_thr=0.45,
+                    min_pos_iou=0.45,
+                    ignore_iof_thr=-1)
+            ],
+            allowed_border=0,
+            pos_weight=-1,
+            debug=False),
+        rpn_proposal=dict(
+            nms_pre=9000,
+            nms_post=512,
+            max_num=512,
+            nms_thr=0.8,
+            score_thr=0,
+            use_rotate_nms=True),
+        rcnn=dict(
+            assigner=[
+                dict(  # for Pedestrian
+                    type='Max3DIoUAssigner',
+                    iou_calculator=dict(
+                        type='BboxOverlaps3D', coordinate='lidar'),
+                    pos_iou_thr=0.55,
+                    neg_iou_thr=0.55,
+                    min_pos_iou=0.55,
+                    ignore_iof_thr=-1),
+                dict(  # for Cyclist
+                    type='Max3DIoUAssigner',
+                    iou_calculator=dict(
+                        type='BboxOverlaps3D', coordinate='lidar'),
+                    pos_iou_thr=0.55,
+                    neg_iou_thr=0.55,
+                    min_pos_iou=0.55,
+                    ignore_iof_thr=-1),
+                dict(  # for Car
+                    type='Max3DIoUAssigner',
+                    iou_calculator=dict(
+                        type='BboxOverlaps3D', coordinate='lidar'),
+                    pos_iou_thr=0.55,
+                    neg_iou_thr=0.55,
+                    min_pos_iou=0.55,
+                    ignore_iof_thr=-1)
+            ],
+            sampler=dict(
+                type='IoUNegPiecewiseSampler',
+                num=128,
+                pos_fraction=0.5,
+                neg_piece_fractions=[0.8, 0.2],
+                neg_iou_piece_thrs=[0.55, 0.1],
+                neg_pos_ub=-1,
+                add_gt_as_proposals=False,
+                return_iou=True),
+            cls_pos_thr=0.75,
+            cls_neg_thr=0.25)),
+    test_cfg=dict(
+        rpn=dict(
+            nms_pre=1024,
+            nms_post=100,
+            max_num=100,
+            nms_thr=0.7,
+            score_thr=0,
+            use_rotate_nms=True),
+        rcnn=dict(
+            use_rotate_nms=True,
+            use_raw_score=True,
+            nms_thr=0.1,
+            score_thr=0.1)))
+train_dataloader = dict(
+    batch_size=2,
+    num_workers=2,
+    dataset=dict(dataset=dict(pipeline=train_pipeline, metainfo=metainfo)))
+test_dataloader = dict(dataset=dict(pipeline=test_pipeline, metainfo=metainfo))
+eval_dataloader = dict(dataset=dict(pipeline=test_pipeline, metainfo=metainfo))
+lr = 0.001
+optim_wrapper = dict(optimizer=dict(lr=lr))
+param_scheduler = [
+    # learning rate scheduler
+    # During the first 16 epochs, learning rate increases from 0 to lr * 10
+    # during the next 24 epochs, learning rate decreases from lr * 10 to
+    # lr * 1e-4
+    dict(
+        type='CosineAnnealingLR',
+        T_max=15,
+        eta_min=lr * 10,
+        begin=0,
+        end=15,
+        by_epoch=True,
+        convert_to_iter_based=True),
+    dict(
+        type='CosineAnnealingLR',
+        T_max=25,
+        eta_min=lr * 1e-4,
+        begin=15,
+        end=40,
+        by_epoch=True,
+        convert_to_iter_based=True),
+    # momentum scheduler
+    # During the first 16 epochs, momentum increases from 0 to 0.85 / 0.95
+    # during the next 24 epochs, momentum increases from 0.85 / 0.95 to 1
+    dict(
+        type='CosineAnnealingMomentum',
+        T_max=15,
+        eta_min=0.85 / 0.95,
+        begin=0,
+        end=15,
+        by_epoch=True,
+        convert_to_iter_based=True),
+    dict(
+        type='CosineAnnealingMomentum',
+        T_max=25,
+        eta_min=1,
+        begin=15,
+        end=40,
+        by_epoch=True,
+        convert_to_iter_based=True)
+]

mmdet3d/apis/inference.py

@@ -67,7 +67,6 @@ def init_model(config: Union[str, Path, Config],
 
     if checkpoint is not None:
         checkpoint = load_checkpoint(model, checkpoint, map_location='cpu')
-
         dataset_meta = checkpoint['meta'].get('dataset_meta', None)
         # save the dataset_meta in the model for convenience
         if 'dataset_meta' in checkpoint.get('meta', {}):

mmdet3d/models/detectors/__init__.py

@@ -14,6 +14,7 @@ from .mvx_faster_rcnn import DynamicMVXFasterRCNN, MVXFasterRCNN
 from .mvx_two_stage import MVXTwoStageDetector
 from .parta2 import PartA2
 from .point_rcnn import PointRCNN
+from .pv_rcnn import PointVoxelRCNN
 from .sassd import SASSD
 from .single_stage_mono3d import SingleStageMono3DDetector
 from .smoke_mono3d import SMOKEMono3D
@@ -26,5 +27,6 @@ __all__ = [
     'DynamicMVXFasterRCNN', 'MVXFasterRCNN', 'PartA2', 'VoteNet', 'H3DNet',
     'CenterPoint', 'SSD3DNet', 'ImVoteNet', 'SingleStageMono3DDetector',
     'FCOSMono3D', 'ImVoxelNet', 'GroupFree3DNet', 'PointRCNN', 'SMOKEMono3D',
-    'SASSD', 'MinkSingleStage3DDetector', 'MultiViewDfM', 'DfM'
+    'SASSD', 'MinkSingleStage3DDetector', 'MultiViewDfM', 'DfM',
+    'PointVoxelRCNN'
 ]

mmdet3d/models/detectors/pv_rcnn.py

+# Copyright (c) OpenMMLab. All rights reserved.
+import copy
+from typing import Optional
+
+from mmdet3d.registry import MODELS
+from mmdet3d.structures.det3d_data_sample import SampleList
+from mmdet3d.utils import InstanceList
+from .two_stage import TwoStage3DDetector
+
+
+@MODELS.register_module()
+class PointVoxelRCNN(TwoStage3DDetector):
+    r"""PointVoxelRCNN detector.
+
+    Please refer to the `PointVoxelRCNN <https://arxiv.org/abs/1912.13192>`_.
+
+    Args:
+        voxel_encoder (dict): Point voxelization encoder layer.
+        middle_encoder (dict): Middle encoder layer
+            of points cloud modality.
+        backbone (dict): Backbone of extracting points features.
+        neck (dict, optional): Neck of extracting points features.
+            Defaults to None.
+        rpn_head (dict, optional): Config of RPN head. Defaults to None.
+        points_encoder (dict, optional): Points encoder to extract point-wise
+            features. Defaults to None.
+        roi_head (dict, optional): Config of ROI head. Defaults to None.
+        train_cfg (dict, optional): Train config of model.
+            Defaults to None.
+        test_cfg (dict, optional): Train config of model.
+            Defaults to None.
+        init_cfg (dict, optional): Initialize config of
+            model. Defaults to None.
+        data_preprocessor (dict or ConfigDict, optional): The pre-process
+            config of :class:`Det3DDataPreprocessor`. Defaults to None.
+    """
+
+    def __init__(self,
+                 voxel_encoder: dict,
+                 middle_encoder: dict,
+                 backbone: dict,
+                 neck: Optional[dict] = None,
+                 rpn_head: Optional[dict] = None,
+                 points_encoder: Optional[dict] = None,
+                 roi_head: Optional[dict] = None,
+                 train_cfg: Optional[dict] = None,
+                 test_cfg: Optional[dict] = None,
+                 init_cfg: Optional[dict] = None,
+                 data_preprocessor: Optional[dict] = None) -> None:
+        super().__init__(
+            backbone=backbone,
+            neck=neck,
+            rpn_head=rpn_head,
+            roi_head=roi_head,
+            train_cfg=train_cfg,
+            test_cfg=test_cfg,
+            init_cfg=init_cfg,
+            data_preprocessor=data_preprocessor)
+        self.voxel_encoder = MODELS.build(voxel_encoder)
+        self.middle_encoder = MODELS.build(middle_encoder)
+        self.points_encoder = MODELS.build(points_encoder)
+
+    def predict(self, batch_inputs_dict: dict, batch_data_samples: SampleList,
+                **kwargs) -> SampleList:
+        """Predict results from a batch of inputs and data samples with post-
+        processing.
+
+        Args:
+            batch_inputs_dict (dict): The model input dict which include
+                'points', 'voxels' keys.
+
+                    - points (list[torch.Tensor]): Point cloud of each sample.
+                    - voxels (dict[torch.Tensor]): Voxels of the batch sample.
+
+            batch_data_samples (List[:obj:`Det3DDataSample`]): The Data
+                samples. It usually includes information such as
+                `gt_instance_3d`, `gt_panoptic_seg_3d` and `gt_sem_seg_3d`.
+
+        Returns:
+            list[:obj:`Det3DDataSample`]: Detection results of the
+            input samples. Each Det3DDataSample usually contain
+            'pred_instances_3d'. And the ``pred_instances_3d`` usually
+            contains following keys.
+
+                - scores_3d (Tensor): Classification scores, has a shape
+                    (num_instance, )
+                - labels_3d (Tensor): Labels of bboxes, has a shape
+                    (num_instances, ).
+                - bboxes_3d (Tensor): Contains a tensor with shape
+                    (num_instances, C) where C >=7.
+        """
+        feats_dict = self.extract_feat(batch_inputs_dict)
+        if self.with_rpn:
+            rpn_results_list = self.rpn_head.predict(feats_dict,
+                                                     batch_data_samples)
+        else:
+            rpn_results_list = [
+                data_sample.proposals for data_sample in batch_data_samples
+            ]
+
+        # extrack points feats by points_encoder
+        points_feats_dict = self.extract_points_feat(batch_inputs_dict,
+                                                     feats_dict,
+                                                     rpn_results_list)
+
+        results_list_3d = self.roi_head.predict(points_feats_dict,
+                                                rpn_results_list,
+                                                batch_data_samples)
+
+        # connvert to Det3DDataSample
+        results_list = self.add_pred_to_datasample(batch_data_samples,
+                                                   results_list_3d)
+
+        return results_list
+
+    def extract_feat(self, batch_inputs_dict: dict) -> dict:
+        """Extract features from the input voxels.
+
+        Args:
+            batch_inputs_dict (dict): The model input dict which include
+                'points', 'voxels' keys.
+
+                - points (list[torch.Tensor]): Point cloud of each sample.
+                - voxels (dict[torch.Tensor]): Voxels of the batch sample.
+
+        Returns:
+            dict: We typically obtain a dict of features from the backbone +
+                neck, it includes:
+
+                - spatial_feats (torch.Tensor): Spatial feats from middle
+                    encoder.
+                - multi_scale_3d_feats (list[torch.Tensor]): Multi scale
+                    middle feats from middle encoder.
+                - neck_feats (torch.Tensor): Neck feats from neck.
+        """
+        feats_dict = dict()
+        voxel_dict = batch_inputs_dict['voxels']
+        voxel_features = self.voxel_encoder(voxel_dict['voxels'],
+                                            voxel_dict['num_points'],
+                                            voxel_dict['coors'])
+        batch_size = voxel_dict['coors'][-1, 0].item() + 1
+        feats_dict['spatial_feats'], feats_dict[
+            'multi_scale_3d_feats'] = self.middle_encoder(
+                voxel_features, voxel_dict['coors'], batch_size)
+        x = self.backbone(feats_dict['spatial_feats'])
+        if self.with_neck:
+            neck_feats = self.neck(x)
+            feats_dict['neck_feats'] = neck_feats
+        return feats_dict
+
+    def extract_points_feat(self, batch_inputs_dict: dict, feats_dict: dict,
+                            rpn_results_list: InstanceList) -> dict:
+        """Extract point-wise features from the raw points and voxel features.
+
+        Args:
+            batch_inputs_dict (dict): The model input dict which include
+                'points', 'voxels' keys.
+
+                - points (list[torch.Tensor]): Point cloud of each sample.
+                - voxels (dict[torch.Tensor]): Voxels of the batch sample.
+            feats_dict (dict): Contains features from the first stage.
+            rpn_results_list (List[:obj:`InstanceData`]): Detection results
+                of rpn head.
+
+        Returns:
+            dict: Contain Point-wise features, include:
+                - keypoints (torch.Tensor): Sampled key points.
+                - keypoint_features (torch.Tensor): Gather key points features
+                    from multi input.
+                - fusion_keypoint_features (torch.Tensor): Fusion
+                    keypoint_features by point_feature_fusion_layer.
+        """
+        return self.points_encoder(batch_inputs_dict, feats_dict,
+                                   rpn_results_list)
+
+    def loss(self, batch_inputs_dict: dict, batch_data_samples: SampleList,
+             **kwargs):
+        """Calculate losses from a batch of inputs and data samples.
+
+        Args:
+            batch_inputs_dict (dict): The model input dict which include
+                'points', 'voxels' keys.
+
+                - points (list[torch.Tensor]): Point cloud of each sample.
+                - voxels (dict[torch.Tensor]): Voxels of the batch sample.
+
+            batch_data_samples (List[:obj:`Det3DDataSample`]): The Data
+                samples. It usually includes information such as
+                `gt_instance_3d`, `gt_panoptic_seg_3d` and `gt_sem_seg_3d`.
+
+        Returns:
+            dict: A dictionary of loss components.
+        """
+        feats_dict = self.extract_feat(batch_inputs_dict)
+
+        losses = dict()
+
+        # RPN forward and loss
+        if self.with_rpn:
+            proposal_cfg = self.train_cfg.get('rpn_proposal',
+                                              self.test_cfg.rpn)
+            rpn_data_samples = copy.deepcopy(batch_data_samples)
+
+            rpn_losses, rpn_results_list = self.rpn_head.loss_and_predict(
+                feats_dict,
+                rpn_data_samples,
+                proposal_cfg=proposal_cfg,
+                **kwargs)
+            # avoid get same name with roi_head loss
+            keys = rpn_losses.keys()
+            for key in keys:
+                if 'loss' in key and 'rpn' not in key:
+                    rpn_losses[f'rpn_{key}'] = rpn_losses.pop(key)
+            losses.update(rpn_losses)
+        else:
+            # TODO: Not support currently, should have a check at Fast R-CNN
+            assert batch_data_samples[0].get('proposals', None) is not None
+            # use pre-defined proposals in InstanceData for the second stage
+            # to extract ROI features.
+            rpn_results_list = [
+                data_sample.proposals for data_sample in batch_data_samples
+            ]
+
+        points_feats_dict = self.extract_points_feat(batch_inputs_dict,
+                                                     feats_dict,
+                                                     rpn_results_list)
+
+        roi_losses = self.roi_head.loss(points_feats_dict, rpn_results_list,
+                                        batch_data_samples)
+        losses.update(roi_losses)
+
+        return losses

mmdet3d/models/layers/pointnet_modules/__init__.py

@@ -4,9 +4,10 @@ from .paconv_sa_module import (PAConvCUDASAModule, PAConvCUDASAModuleMSG,
                                PAConvSAModule, PAConvSAModuleMSG)
 from .point_fp_module import PointFPModule
 from .point_sa_module import PointSAModule, PointSAModuleMSG
+from .stack_point_sa_module import StackedSAModuleMSG
 
 __all__ = [
     'build_sa_module', 'PointSAModuleMSG', 'PointSAModule', 'PointFPModule',
     'PAConvSAModule', 'PAConvSAModuleMSG', 'PAConvCUDASAModule',
-    'PAConvCUDASAModuleMSG'
+    'PAConvCUDASAModuleMSG', 'StackedSAModuleMSG'
 ]

mmdet3d/models/layers/pointnet_modules/stack_point_sa_module.py

+# Copyright (c) OpenMMLab. All rights reserved.
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from mmcv.cnn import ConvModule
+from mmcv.ops import ball_query, grouping_operation
+from mmengine.model import BaseModule
+from torch import Tensor
+
+from mmdet3d.registry import MODELS
+
+
+class StackQueryAndGroup(BaseModule):
+    """Find nearby points in spherical space.
+
+    Args:
+        radius (float): List of radius in each ball query.
+        sample_nums (int): Number of samples in each ball query.
+        use_xyz (bool): Whether to use xyz. Default: True.
+        init_cfg (dict, optional): Initialize config of
+            model. Defaults to None.
+    """
+
+    def __init__(self,
+                 radius: float,
+                 sample_nums: int,
+                 use_xyz: bool = True,
+                 init_cfg: dict = None):
+        super().__init__(init_cfg=init_cfg)
+        self.radius, self.sample_nums, self.use_xyz = \
+            radius, sample_nums, use_xyz
+
+    def forward(self,
+                xyz: torch.Tensor,
+                xyz_batch_cnt: torch.Tensor,
+                new_xyz: torch.Tensor,
+                new_xyz_batch_cnt: torch.Tensor,
+                features: torch.Tensor = None) -> Tuple[Tensor, Tensor]:
+        """Forward.
+
+        Args:
+            xyz (Tensor): Tensor of the xyz coordinates
+                of the features shape with (N1 + N2 ..., 3).
+            xyz_batch_cnt: (Tensor): Stacked input xyz coordinates nums in
+                each batch, just like (N1, N2, ...).
+            new_xyz (Tensor): New coords of the outputs shape with
+                (M1 + M2 ..., 3).
+            new_xyz_batch_cnt: (Tensor): Stacked new xyz coordinates nums
+                in each batch, just like (M1, M2, ...).
+            features (Tensor, optional): Features of each point with shape
+                (N1 + N2 ..., C). C is features channel number. Default: None.
+        """
+        assert xyz.shape[0] == xyz_batch_cnt.sum(
+        ), f'xyz: {str(xyz.shape)}, xyz_batch_cnt: str(new_xyz_batch_cnt)'
+        assert new_xyz.shape[0] == new_xyz_batch_cnt.sum(), \
+            'new_xyz: str(new_xyz.shape), new_xyz_batch_cnt: ' \
+            'str(new_xyz_batch_cnt)'
+
+        # idx: (M1 + M2 ..., nsample), empty_ball_mask: (M1 + M2 ...)
+        idx, empty_ball_mask = ball_query(0, self.radius, self.sample_nums,
+                                          xyz, new_xyz, xyz_batch_cnt,
+                                          new_xyz_batch_cnt)
+        grouped_xyz = grouping_operation(
+            xyz, idx, xyz_batch_cnt,
+            new_xyz_batch_cnt)  # (M1 + M2, 3, nsample)
+        grouped_xyz -= new_xyz.unsqueeze(-1)
+
+        grouped_xyz[empty_ball_mask] = 0
+        if features is not None:
+            grouped_features = grouping_operation(
+                features, idx, xyz_batch_cnt,
+                new_xyz_batch_cnt)  # (M1 + M2, C, nsample)
+            grouped_features[empty_ball_mask] = 0
+            if self.use_xyz:
+                new_features = torch.cat(
+                    [grouped_xyz, grouped_features],
+                    dim=1)  # (M1 + M2 ..., C + 3, nsample)
+            else:
+                new_features = grouped_features
+        else:
+            assert self.use_xyz, 'Cannot have not features and not' \
+                                 ' use xyz as a feature!'
+            new_features = grouped_xyz
+        return new_features, idx
+
+
+@MODELS.register_module()
+class StackedSAModuleMSG(BaseModule):
+    """Stack point set abstraction module.
+
+    Args:
+        in_channels (int): Input channels.
+        radius (list[float]): List of radius in each ball query.
+        sample_nums (list[int]): Number of samples in each ball query.
+        mlp_channels (list[list[int]]): Specify mlp channels of the
+            pointnet before the global pooling for each scale to encode
+            point features.
+        use_xyz (bool): Whether to use xyz. Default: True.
+        pool_mod (str): Type of pooling method.
+            Default: 'max_pool'.
+        norm_cfg (dict): Type of normalization method. Defaults to
+            dict(type='BN2d', eps=1e-5, momentum=0.01).
+        init_cfg (dict, optional): Initialize config of
+            model. Defaults to None.
+    """
+
+    def __init__(self,
+                 in_channels: int,
+                 radius: List[float],
+                 sample_nums: List[int],
+                 mlp_channels: List[List[int]],
+                 use_xyz: bool = True,
+                 pool_mod='max',
+                 norm_cfg: dict = dict(type='BN2d', eps=1e-5, momentum=0.01),
+                 init_cfg: dict = None,
+                 **kwargs) -> None:
+        super(StackedSAModuleMSG, self).__init__(init_cfg=init_cfg)
+        assert len(radius) == len(sample_nums) == len(mlp_channels)
+
+        self.groupers = nn.ModuleList()
+        self.mlps = nn.ModuleList()
+        for i in range(len(radius)):
+            cin = in_channels
+            if use_xyz:
+                cin += 3
+            cur_radius = radius[i]
+            nsample = sample_nums[i]
+            mlp_spec = mlp_channels[i]
+
+            self.groupers.append(
+                StackQueryAndGroup(cur_radius, nsample, use_xyz=use_xyz))
+
+            mlp = nn.Sequential()
+            for i in range(len(mlp_spec)):
+                cout = mlp_spec[i]
+                mlp.add_module(
+                    f'layer{i}',
+                    ConvModule(
+                        cin,
+                        cout,
+                        kernel_size=(1, 1),
+                        stride=(1, 1),
+                        conv_cfg=dict(type='Conv2d'),
+                        norm_cfg=norm_cfg,
+                        bias=False))
+                cin = cout
+            self.mlps.append(mlp)
+        self.pool_mod = pool_mod
+
+    def forward(self,
+                xyz: Tensor,
+                xyz_batch_cnt: Tensor,
+                new_xyz: Tensor,
+                new_xyz_batch_cnt: Tensor,
+                features: Optional[Tensor] = None) -> Tuple[Tensor, Tensor]:
+        """Forward.
+
+        Args:
+            xyz (Tensor): Tensor of the xyz coordinates
+                of the features shape with (N1 + N2 ..., 3).
+            xyz_batch_cnt: (Tensor): Stacked input xyz coordinates nums in
+                each batch, just like (N1, N2, ...).
+            new_xyz (Tensor): New coords of the outputs shape with
+                (M1 + M2 ..., 3).
+            new_xyz_batch_cnt: (Tensor): Stacked new xyz coordinates nums
+                in each batch, just like (M1, M2, ...).
+            features (Tensor, optional): Features of each point with shape
+                (N1 + N2 ..., C). C is features channel number. Default: None.
+
+        Returns:
+            Return new points coordinates and features:
+                - new_xyz  (Tensor): Target points coordinates with shape
+                    (N1 + N2 ..., 3).
+                - new_features (Tensor): Target points features with shape
+                    (M1 + M2 ..., sum_k(mlps[k][-1])).
+        """
+        new_features_list = []
+        for k in range(len(self.groupers)):
+            grouped_features, ball_idxs = self.groupers[k](
+                xyz, xyz_batch_cnt, new_xyz, new_xyz_batch_cnt,
+                features)  # (M1 + M2, Cin, nsample)
+            grouped_features = grouped_features.permute(1, 0,
+                                                        2).unsqueeze(dim=0)
+            new_features = self.mlps[k](grouped_features)
+            # (M1 + M2 ..., Cout, nsample)
+            if self.pool_mod == 'max':
+                new_features = new_features.max(-1).values
+            elif self.pool_mod == 'avg':
+                new_features = new_features.mean(-1)
+            else:
+                raise NotImplementedError
+            new_features = new_features.squeeze(dim=0).permute(1, 0)
+            new_features_list.append(new_features)
+
+        new_features = torch.cat(new_features_list, dim=1)
+
+        return new_xyz, new_features

mmdet3d/models/middle_encoders/__init__.py

@@ -2,7 +2,9 @@
 from .pillar_scatter import PointPillarsScatter
 from .sparse_encoder import SparseEncoder, SparseEncoderSASSD
 from .sparse_unet import SparseUNet
+from .voxel_set_abstraction import VoxelSetAbstraction
 
 __all__ = [
-    'PointPillarsScatter', 'SparseEncoder', 'SparseEncoderSASSD', 'SparseUNet'
+    'PointPillarsScatter', 'SparseEncoder', 'SparseEncoderSASSD', 'SparseUNet',
+    'VoxelSetAbstraction'
 ]

mmdet3d/models/middle_encoders/sparse_encoder.py

@@ -41,6 +41,8 @@ class SparseEncoder(nn.Module):
             Defaults to ((1, ), (1, 1, 1), (1, 1, 1), ((0, 1, 1), 1, 1)).
         block_type (str, optional): Type of the block to use.
             Defaults to 'conv_module'.
+        return_middle_feats (bool): Whether output middle features.
+            Default to False.
     """
 
     def __init__(self,
@@ -54,7 +56,8 @@ class SparseEncoder(nn.Module):
                                                                         64)),
                  encoder_paddings=((1, ), (1, 1, 1), (1, 1, 1), ((0, 1, 1), 1,
                                                                  1)),
-                 block_type='conv_module'):
+                 block_type='conv_module',
+                 return_middle_feats=False):
         super().__init__()
         assert block_type in ['conv_module', 'basicblock']
         self.sparse_shape = sparse_shape
@@ -66,6 +69,7 @@ class SparseEncoder(nn.Module):
         self.encoder_paddings = encoder_paddings
         self.stage_num = len(self.encoder_channels)
         self.fp16_enabled = False
+        self.return_middle_feats = return_middle_feats
         # Spconv init all weight on its own
 
         assert isinstance(order, tuple) and len(order) == 3
@@ -117,7 +121,14 @@ class SparseEncoder(nn.Module):
             batch_size (int): Batch size.
 
         Returns:
-            dict: Backbone features.
+            torch.Tensor | tuple[torch.Tensor, list]: Return spatial features
+                include:
+
+            - spatial_features (torch.Tensor): Spatial features are out from
+                the last layer.
+            - encode_features (List[SparseConvTensor], optional): Middle layer
+                output features. When self.return_middle_feats is True, the
+                module returns middle features.
         """
         coors = coors.int()
         input_sp_tensor = SparseConvTensor(voxel_features, coors,
@@ -137,7 +148,10 @@ class SparseEncoder(nn.Module):
         N, C, D, H, W = spatial_features.shape
         spatial_features = spatial_features.view(N, C * D, H, W)
 
-        return spatial_features
+        if self.return_middle_feats:
+            return spatial_features, encode_features
+        else:
+            return spatial_features
 
     def make_encoder_layers(self,
                             make_block,

mmdet3d/models/middle_encoders/voxel_set_abstraction.py

+# Copyright (c) OpenMMLab. All rights reserved.
+from typing import List, Optional
+
+import mmengine
+import torch
+import torch.nn as nn
+from mmcv.cnn import ConvModule
+from mmcv.ops.furthest_point_sample import furthest_point_sample
+from mmengine.model import BaseModule
+
+from mmdet3d.registry import MODELS
+from mmdet3d.utils import InstanceList
+
+
+def bilinear_interpolate_torch(inputs, x, y):
+    """Bilinear interpolate for inputs."""
+    x0 = torch.floor(x).long()
+    x1 = x0 + 1
+
+    y0 = torch.floor(y).long()
+    y1 = y0 + 1
+
+    x0 = torch.clamp(x0, 0, inputs.shape[1] - 1)
+    x1 = torch.clamp(x1, 0, inputs.shape[1] - 1)
+    y0 = torch.clamp(y0, 0, inputs.shape[0] - 1)
+    y1 = torch.clamp(y1, 0, inputs.shape[0] - 1)
+
+    Ia = inputs[y0, x0]
+    Ib = inputs[y1, x0]
+    Ic = inputs[y0, x1]
+    Id = inputs[y1, x1]
+
+    wa = (x1.type_as(x) - x) * (y1.type_as(y) - y)
+    wb = (x1.type_as(x) - x) * (y - y0.type_as(y))
+    wc = (x - x0.type_as(x)) * (y1.type_as(y) - y)
+    wd = (x - x0.type_as(x)) * (y - y0.type_as(y))
+    ans = torch.t((torch.t(Ia) * wa)) + torch.t(torch.t(Ib) * wb) + torch.t(
+        torch.t(Ic) * wc) + torch.t(torch.t(Id) * wd)
+    return ans
+
+
+@MODELS.register_module()
+class VoxelSetAbstraction(BaseModule):
+    """Voxel set abstraction module for PVRCNN and PVRCNN++.
+
+    Args:
+        num_keypoints (int): The number of key points sampled from
+            raw points cloud.
+        fused_out_channel (int): Key points feature output channels
+            num after fused. Default to 128.
+        voxel_size (list[float]): Size of voxels. Defaults to
+            [0.05, 0.05, 0.1].
+        point_cloud_range (list[float]): Point cloud range. Defaults to
+            [0, -40, -3, 70.4, 40, 1].
+        voxel_sa_cfgs_list (List[dict or ConfigDict], optional): List of SA
+            module cfg. Used to gather key points features from multi-wise
+            voxel features. Default to None.
+        rawpoints_sa_cfgs (dict or ConfigDict, optional): SA module cfg.
+            Used to gather key points features from raw points. Default to
+            None.
+        bev_feat_channel (int): Bev features channels num.
+            Default to 256.
+        bev_scale_factor (int): Bev features scale factor. Default to 8.
+        voxel_center_as_source (bool): Whether used voxel centers as points
+            cloud key points. Defaults to False.
+        norm_cfg (dict[str]): Config of normalization layer. Default
+            used dict(type='BN1d', eps=1e-5, momentum=0.1).
+        bias (bool | str, optional): If specified as `auto`, it will be
+            decided by `norm_cfg`. `bias` will be set as True if
+            `norm_cfg` is None, otherwise False. Default: 'auto'.
+    """
+
+    def __init__(self,
+                 num_keypoints: int,
+                 fused_out_channel: int = 128,
+                 voxel_size: list = [0.05, 0.05, 0.1],
+                 point_cloud_range: list = [0, -40, -3, 70.4, 40, 1],
+                 voxel_sa_cfgs_list: Optional[list] = None,
+                 rawpoints_sa_cfgs: Optional[dict] = None,
+                 bev_feat_channel: int = 256,
+                 bev_scale_factor: int = 8,
+                 voxel_center_as_source: bool = False,
+                 norm_cfg: dict = dict(type='BN2d', eps=1e-5, momentum=0.1),
+                 bias: str = 'auto') -> None:
+        super().__init__()
+        self.num_keypoints = num_keypoints
+        self.fused_out_channel = fused_out_channel
+        self.voxel_size = voxel_size
+        self.point_cloud_range = point_cloud_range
+        self.voxel_center_as_source = voxel_center_as_source
+
+        gathered_channel = 0
+
+        if rawpoints_sa_cfgs is not None:
+            self.rawpoints_sa_layer = MODELS.build(rawpoints_sa_cfgs)
+            gathered_channel += sum(
+                [x[-1] for x in rawpoints_sa_cfgs.mlp_channels])
+        else:
+            self.rawpoints_sa_layer = None
+
+        if voxel_sa_cfgs_list is not None:
+            self.voxel_sa_configs_list = voxel_sa_cfgs_list
+            self.voxel_sa_layers = nn.ModuleList()
+            for voxel_sa_config in voxel_sa_cfgs_list:
+                cur_layer = MODELS.build(voxel_sa_config)
+                self.voxel_sa_layers.append(cur_layer)
+                gathered_channel += sum(
+                    [x[-1] for x in voxel_sa_config.mlp_channels])
+        else:
+            self.voxel_sa_layers = None
+
+        if bev_feat_channel is not None and bev_scale_factor is not None:
+            self.bev_cfg = mmengine.Config(
+                dict(
+                    bev_feat_channels=bev_feat_channel,
+                    bev_scale_factor=bev_scale_factor))
+            gathered_channel += bev_feat_channel
+        else:
+            self.bev_cfg = None
+        self.point_feature_fusion_layer = nn.Sequential(
+            ConvModule(
+                gathered_channel,
+                fused_out_channel,
+                kernel_size=(1, 1),
+                stride=(1, 1),
+                conv_cfg=dict(type='Conv2d'),
+                norm_cfg=norm_cfg,
+                bias=bias))
+
+    def interpolate_from_bev_features(self, keypoints: torch.Tensor,
+                                      bev_features: torch.Tensor,
+                                      batch_size: int,
+                                      bev_scale_factor: int) -> torch.Tensor:
+        """Gather key points features from bev feature map by interpolate.
+
+        Args:
+            keypoints (torch.Tensor): Sampled key points with shape
+                (N1 + N2 + ..., NDim).
+            bev_features (torch.Tensor): Bev feature map from the first
+                stage with shape (B, C, H, W).
+            batch_size (int): Input batch size.
+            bev_scale_factor (int): Bev feature map scale factor.
+
+        Returns:
+            torch.Tensor: Key points features gather from bev feature
+                map with shape (N1 + N2 + ..., C)
+        """
+        x_idxs = (keypoints[..., 0] -
+                  self.point_cloud_range[0]) / self.voxel_size[0]
+        y_idxs = (keypoints[..., 1] -
+                  self.point_cloud_range[1]) / self.voxel_size[1]
+
+        x_idxs = x_idxs / bev_scale_factor
+        y_idxs = y_idxs / bev_scale_factor
+
+        point_bev_features_list = []
+        for k in range(batch_size):
+            cur_x_idxs = x_idxs[k, ...]
+            cur_y_idxs = y_idxs[k, ...]
+            cur_bev_features = bev_features[k].permute(1, 2, 0)  # (H, W, C)
+            point_bev_features = bilinear_interpolate_torch(
+                cur_bev_features, cur_x_idxs, cur_y_idxs)
+            point_bev_features_list.append(point_bev_features)
+
+        point_bev_features = torch.cat(
+            point_bev_features_list, dim=0)  # (N1 + N2 + ..., C)
+        return point_bev_features.view(batch_size, keypoints.shape[1], -1)
+
+    def get_voxel_centers(self, coors: torch.Tensor,
+                          scale_factor: float) -> torch.Tensor:
+        """Get voxel centers coordinate.
+
+        Args:
+            coors (torch.Tensor): Coordinates of voxels shape is Nx(1+NDim),
+                where 1 represents the batch index.
+            scale_factor (float): Scale factor.
+
+        Returns:
+            torch.Tensor: Voxel centers coordinate with shape (N, 3).
+        """
+        assert coors.shape[1] == 4
+        voxel_centers = coors[:, [3, 2, 1]].float()  # (xyz)
+        voxel_size = torch.tensor(
+            self.voxel_size,
+            device=voxel_centers.device).float() * scale_factor
+        pc_range = torch.tensor(
+            self.point_cloud_range[0:3], device=voxel_centers.device).float()
+        voxel_centers = (voxel_centers + 0.5) * voxel_size + pc_range
+        return voxel_centers
+
+    def sample_key_points(self, points: List[torch.Tensor],
+                          coors: torch.Tensor) -> torch.Tensor:
+        """Sample key points from raw points cloud.
+
+        Args:
+            points (List[torch.Tensor]): Point cloud of each sample.
+            coors (torch.Tensor): Coordinates of voxels shape is Nx(1+NDim),
+                where 1 represents the batch index.
+
+        Returns:
+            torch.Tensor: (B, M, 3) Key points of each sample.
+                M is num_keypoints.
+        """
+        assert points is not None or coors is not None
+        if self.voxel_center_as_source:
+            _src_points = self.get_voxel_centers(coors=coors, scale_factor=1)
+            batch_size = coors[-1, 0].item() + 1
+            src_points = [
+                _src_points[coors[:, 0] == b] for b in range(batch_size)
+            ]
+        else:
+            src_points = [p[..., :3] for p in points]
+
+        keypoints_list = []
+        for points_to_sample in src_points:
+            num_points = points_to_sample.shape[0]
+            cur_pt_idxs = furthest_point_sample(
+                points_to_sample.unsqueeze(dim=0).contiguous(),
+                self.num_keypoints).long()[0]
+
+            if num_points < self.num_keypoints:
+                times = int(self.num_keypoints / num_points) + 1
+                non_empty = cur_pt_idxs[:num_points]
+                cur_pt_idxs = non_empty.repeat(times)[:self.num_keypoints]
+
+            keypoints = points_to_sample[cur_pt_idxs]
+
+            keypoints_list.append(keypoints)
+        keypoints = torch.stack(keypoints_list, dim=0)  # (B, M, 3)
+        return keypoints
+
+    def forward(self, batch_inputs_dict: dict, feats_dict: dict,
+                rpn_results_list: InstanceList) -> dict:
+        """Extract point-wise features from multi-input.
+
+        Args:
+            batch_inputs_dict (dict): The model input dict which include
+                'points', 'voxels' keys.
+
+                - points (list[torch.Tensor]): Point cloud of each sample.
+                - voxels (dict[torch.Tensor]): Voxels of the batch sample.
+            feats_dict (dict): Contains features from the first
+                stage.
+            rpn_results_list (List[:obj:`InstanceData`]): Detection results
+                of rpn head.
+
+        Returns:
+            dict: Contain Point-wise features, include:
+                - keypoints (torch.Tensor): Sampled key points.
+                - keypoint_features (torch.Tensor): Gathered key points
+                    features from multi input.
+                - fusion_keypoint_features (torch.Tensor): Fusion
+                    keypoint_features by point_feature_fusion_layer.
+        """
+        points = batch_inputs_dict['points']
+        voxel_encode_features = feats_dict['multi_scale_3d_feats']
+        bev_encode_features = feats_dict['spatial_feats']
+        if self.voxel_center_as_source:
+            voxels_coors = batch_inputs_dict['voxels']['coors']
+        else:
+            voxels_coors = None
+        keypoints = self.sample_key_points(points, voxels_coors)
+
+        point_features_list = []
+        batch_size = len(points)
+
+        if self.bev_cfg is not None:
+            point_bev_features = self.interpolate_from_bev_features(
+                keypoints, bev_encode_features, batch_size,
+                self.bev_cfg.bev_scale_factor)
+            point_features_list.append(point_bev_features.contiguous())
+
+        batch_size, num_keypoints, _ = keypoints.shape
+        key_xyz = keypoints.view(-1, 3)
+        key_xyz_batch_cnt = key_xyz.new_zeros(batch_size).int().fill_(
+            num_keypoints)
+
+        if self.rawpoints_sa_layer is not None:
+            batch_points = torch.cat(points, dim=0)
+            batch_cnt = [len(p) for p in points]
+            xyz = batch_points[:, :3].contiguous()
+            features = None
+            if batch_points.size(1) > 0:
+                features = batch_points[:, 3:].contiguous()
+            xyz_batch_cnt = xyz.new_tensor(batch_cnt, dtype=torch.int32)
+
+            pooled_points, pooled_features = self.rawpoints_sa_layer(
+                xyz=xyz.contiguous(),
+                xyz_batch_cnt=xyz_batch_cnt,
+                new_xyz=key_xyz.contiguous(),
+                new_xyz_batch_cnt=key_xyz_batch_cnt,
+                features=features.contiguous(),
+            )
+
+            point_features_list.append(pooled_features.contiguous().view(
+                batch_size, num_keypoints, -1))
+        if self.voxel_sa_layers is not None:
+            for k, voxel_sa_layer in enumerate(self.voxel_sa_layers):
+                cur_coords = voxel_encode_features[k].indices
+                xyz = self.get_voxel_centers(
+                    coors=cur_coords,
+                    scale_factor=self.voxel_sa_configs_list[k].scale_factor
+                ).contiguous()
+                xyz_batch_cnt = xyz.new_zeros(batch_size).int()
+                for bs_idx in range(batch_size):
+                    xyz_batch_cnt[bs_idx] = (cur_coords[:, 0] == bs_idx).sum()
+
+                pooled_points, pooled_features = voxel_sa_layer(
+                    xyz=xyz.contiguous(),
+                    xyz_batch_cnt=xyz_batch_cnt,
+                    new_xyz=key_xyz.contiguous(),
+                    new_xyz_batch_cnt=key_xyz_batch_cnt,
+                    features=voxel_encode_features[k].features.contiguous(),
+                )
+                point_features_list.append(pooled_features.contiguous().view(
+                    batch_size, num_keypoints, -1))
+
+        point_features = torch.cat(
+            point_features_list, dim=-1).view(batch_size * num_keypoints, -1,
+                                              1)
+
+        fusion_point_features = self.point_feature_fusion_layer(
+            point_features.unsqueeze(dim=-1)).squeeze(dim=-1)
+
+        batch_idxs = torch.arange(
+            batch_size * num_keypoints, device=keypoints.device
+        ) // num_keypoints  # batch indexes of each key points
+        batch_keypoints_xyz = torch.cat(
+            (batch_idxs.to(key_xyz.dtype).unsqueeze(dim=-1), key_xyz), dim=-1)
+
+        return dict(
+            keypoint_features=point_features.squeeze(dim=-1),
+            fusion_keypoint_features=fusion_point_features.squeeze(dim=-1),
+            keypoints=batch_keypoints_xyz)

mmdet3d/models/roi_heads/__init__.py

@@ -5,10 +5,11 @@ from .h3d_roi_head import H3DRoIHead
 from .mask_heads import PointwiseSemanticHead, PrimitiveHead
 from .part_aggregation_roi_head import PartAggregationROIHead
 from .point_rcnn_roi_head import PointRCNNRoIHead
+from .pv_rcnn_roi_head import PVRCNNRoiHead
 from .roi_extractors import Single3DRoIAwareExtractor, SingleRoIExtractor
 
 __all__ = [
     'Base3DRoIHead', 'PartAggregationROIHead', 'PointwiseSemanticHead',
     'Single3DRoIAwareExtractor', 'PartA2BboxHead', 'SingleRoIExtractor',
-    'H3DRoIHead', 'PrimitiveHead', 'PointRCNNRoIHead'
+    'H3DRoIHead', 'PrimitiveHead', 'PointRCNNRoIHead', 'PVRCNNRoiHead'
 ]

mmdet3d/models/roi_heads/bbox_heads/__init__.py

@@ -7,9 +7,10 @@ from mmdet.models.roi_heads.bbox_heads import (BBoxHead, ConvFCBBoxHead,
 from .h3d_bbox_head import H3DBboxHead
 from .parta2_bbox_head import PartA2BboxHead
 from .point_rcnn_bbox_head import PointRCNNBboxHead
+from .pv_rcnn_bbox_head import PVRCNNBBoxHead
 
 __all__ = [
     'BBoxHead', 'ConvFCBBoxHead', 'Shared2FCBBoxHead',
     'Shared4Conv1FCBBoxHead', 'DoubleConvFCBBoxHead', 'PartA2BboxHead',
-    'H3DBboxHead', 'PointRCNNBboxHead'
+    'H3DBboxHead', 'PointRCNNBboxHead', 'PVRCNNBBoxHead'
 ]

mmdet3d/models/roi_heads/bbox_heads/pv_rcnn_bbox_head.py

+# Copyright (c) OpenMMLab. All rights reserved.
+from typing import Dict, List, Optional, Tuple
+
+import numpy as np
+import torch
+from mmcv.cnn import ConvModule
+from mmengine.model import BaseModule
+from mmengine.structures import InstanceData
+from torch import nn as nn
+
+from mmdet3d.models.builder import build_loss
+from mmdet3d.models.layers import nms_bev, nms_normal_bev
+from mmdet3d.registry import MODELS, TASK_UTILS
+from mmdet3d.structures.bbox_3d import (LiDARInstance3DBoxes,
+                                        rotation_3d_in_axis, xywhr2xyxyr)
+from mmdet3d.utils import InstanceList
+from mmdet.models.task_modules.samplers import SamplingResult
+from mmdet.models.utils import multi_apply
+
+
+@MODELS.register_module()
+class PVRCNNBBoxHead(BaseModule):
+    """PVRCNN BBox head.
+
+    Args:
+        in_channels (int): The number of input channel.
+        grid_size (int): The number of grid points in roi bbox.
+        num_classes (int): The number of classes.
+        class_agnostic (bool): Whether generate class agnostic prediction.
+            Defaults to True.
+        shared_fc_channels (tuple(int)): Out channels of each shared fc layer.
+            Defaults to (256, 256).
+        cls_channels (tuple(int)): Out channels of each classification layer.
+            Defaults to (256, 256).
+        reg_channels (tuple(int)): Out channels of each regression layer.
+            Defaults to (256, 256).
+        dropout_ratio (float): Ratio of dropout layer. Defaults to 0.5.
+        with_corner_loss (bool): Whether to use corner loss or not.
+            Defaults to True.
+        bbox_coder (:obj:`BaseBBoxCoder`): Bbox coder for box head.
+            Defaults to dict(type='DeltaXYZWLHRBBoxCoder').
+        norm_cfg (dict): Type of normalization method.
+            Defaults to dict(type='BN1d', eps=1e-5, momentum=0.1)
+        loss_bbox (dict): Config dict of box regression loss.
+        loss_cls (dict): Config dict of classifacation loss.
+        init_cfg (dict, optional): Initialize config of
+            model.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        grid_size: int,
+        num_classes: int,
+        class_agnostic: bool = True,
+        shared_fc_channels: Tuple[int] = (256, 256),
+        cls_channels: Tuple[int] = (256, 256),
+        reg_channels: Tuple[int] = (256, 256),
+        dropout_ratio: float = 0.3,
+        with_corner_loss: bool = True,
+        bbox_coder: dict = dict(type='DeltaXYZWLHRBBoxCoder'),
+        norm_cfg: dict = dict(type='BN2d', eps=1e-5, momentum=0.1),
+        loss_bbox: dict = dict(
+            type='mmdet.SmoothL1Loss', beta=1.0 / 9.0, loss_weight=2.0),
+        loss_cls: dict = dict(
+            type='mmdet.CrossEntropyLoss',
+            use_sigmoid=True,
+            reduction='none',
+            loss_weight=1.0),
+        init_cfg: Optional[dict] = dict(
+            type='Xavier', layer=['Conv2d', 'Conv1d'], distribution='uniform')
+    ) -> None:
+        super(PVRCNNBBoxHead, self).__init__(init_cfg=init_cfg)
+        self.init_cfg = init_cfg
+        self.num_classes = num_classes
+        self.with_corner_loss = with_corner_loss
+        self.class_agnostic = class_agnostic
+        self.bbox_coder = TASK_UTILS.build(bbox_coder)
+        self.loss_bbox = build_loss(loss_bbox)
+        self.loss_cls = build_loss(loss_cls)
+        self.use_sigmoid_cls = loss_cls.get('use_sigmoid', False)
+
+        cls_out_channels = 1 if class_agnostic else num_classes
+        self.reg_out_channels = self.bbox_coder.code_size * cls_out_channels
+        if self.use_sigmoid_cls:
+            self.cls_out_channels = cls_out_channels
+        else:
+            self.cls_out_channels = cls_out_channels + 1
+
+        self.dropout_ratio = dropout_ratio
+        self.grid_size = grid_size
+
+        # PVRCNNBBoxHead model in_channels is num of grid points in roi box.
+        in_channels *= (self.grid_size**3)
+
+        self.in_channels = in_channels
+
+        self.shared_fc_layer = self._make_fc_layers(
+            in_channels, shared_fc_channels,
+            range(len(shared_fc_channels) - 1), norm_cfg)
+        self.cls_layer = self._make_fc_layers(
+            shared_fc_channels[-1],
+            cls_channels,
+            range(1),
+            norm_cfg,
+            out_channels=self.cls_out_channels)
+        self.reg_layer = self._make_fc_layers(
+            shared_fc_channels[-1],
+            reg_channels,
+            range(1),
+            norm_cfg,
+            out_channels=self.reg_out_channels)
+
+    def _make_fc_layers(self,
+                        in_channels: int,
+                        fc_channels: list,
+                        dropout_indices: list,
+                        norm_cfg: dict,
+                        out_channels: Optional[int] = None) -> torch.nn.Module:
+        """Initial a full connection layer.
+
+        Args:
+            in_channels (int): Module in channels.
+            fc_channels (list): Full connection layer channels.
+            dropout_indices (list): Dropout indices.
+            norm_cfg (dict): Type of normalization method.
+            out_channels (int, optional): Module out channels.
+        """
+        fc_layers = []
+        pre_channel = in_channels
+        for k in range(len(fc_channels)):
+            fc_layers.append(
+                ConvModule(
+                    pre_channel,
+                    fc_channels[k],
+                    kernel_size=(1, 1),
+                    stride=(1, 1),
+                    norm_cfg=norm_cfg,
+                    conv_cfg=dict(type='Conv2d'),
+                    bias=False,
+                    inplace=True))
+            pre_channel = fc_channels[k]
+            if self.dropout_ratio >= 0 and k in dropout_indices:
+                fc_layers.append(nn.Dropout(self.dropout_ratio))
+        if out_channels is not None:
+            fc_layers.append(
+                nn.Conv2d(fc_channels[-1], out_channels, 1, bias=True))
+        fc_layers = nn.Sequential(*fc_layers)
+        return fc_layers
+
+    def forward(self, feats: torch.Tensor) -> Tuple[torch.Tensor]:
+        """Forward pvrcnn bbox head.
+
+        Args:
+            feats (torch.Tensor): Batch point-wise features.
+
+        Returns:
+            tuple[torch.Tensor]: Score of class and bbox predictions.
+        """
+        # (B * N, 6, 6, 6, C)
+        rcnn_batch_size = feats.shape[0]
+        feats = feats.permute(0, 4, 1, 2,
+                              3).contiguous().view(rcnn_batch_size, -1, 1, 1)
+        # (BxN, C*6*6*6)
+        shared_feats = self.shared_fc_layer(feats)
+        cls_score = self.cls_layer(shared_feats).transpose(
+            1, 2).contiguous().view(-1, self.cls_out_channels)  # (B, 1)
+        bbox_pred = self.reg_layer(shared_feats).transpose(
+            1, 2).contiguous().view(-1, self.reg_out_channels)  # (B, C)
+        return cls_score, bbox_pred
+
+    def loss(self, cls_score: torch.Tensor, bbox_pred: torch.Tensor,
+             rois: torch.Tensor, labels: torch.Tensor,
+             bbox_targets: torch.Tensor, pos_gt_bboxes: torch.Tensor,
+             reg_mask: torch.Tensor, label_weights: torch.Tensor,
+             bbox_weights: torch.Tensor) -> Dict:
+        """Coumputing losses.
+
+        Args:
+            cls_score (torch.Tensor): Scores of each roi.
+            bbox_pred (torch.Tensor): Predictions of bboxes.
+            rois (torch.Tensor): Roi bboxes.
+            labels (torch.Tensor): Labels of class.
+            bbox_targets (torch.Tensor): Target of positive bboxes.
+            pos_gt_bboxes (torch.Tensor): Ground truths of positive bboxes.
+            reg_mask (torch.Tensor): Mask for positive bboxes.
+            label_weights (torch.Tensor): Weights of class loss.
+            bbox_weights (torch.Tensor): Weights of bbox loss.
+
+        Returns:
+             dict: Computed losses.
+
+             - loss_cls (torch.Tensor): Loss of classes.
+             - loss_bbox (torch.Tensor): Loss of bboxes.
+             - loss_corner (torch.Tensor): Loss of corners.
+        """
+        losses = dict()
+        rcnn_batch_size = cls_score.shape[0]
+
+        # calculate class loss
+        cls_flat = cls_score.view(-1)
+        loss_cls = self.loss_cls(cls_flat, labels, label_weights)
+        losses['loss_cls'] = loss_cls
+
+        # calculate regression loss
+        code_size = self.bbox_coder.code_size
+        pos_inds = (reg_mask > 0)
+        if pos_inds.any() == 0:
+            # fake a part loss
+            losses['loss_bbox'] = 0 * bbox_pred.sum()
+            if self.with_corner_loss:
+                losses['loss_corner'] = 0 * bbox_pred.sum()
+        else:
+            pos_bbox_pred = bbox_pred.view(rcnn_batch_size, -1)[pos_inds]
+            bbox_weights_flat = bbox_weights[pos_inds].view(-1, 1).repeat(
+                1, pos_bbox_pred.shape[-1])
+            loss_bbox = self.loss_bbox(
+                pos_bbox_pred.unsqueeze(dim=0), bbox_targets.unsqueeze(dim=0),
+                bbox_weights_flat.unsqueeze(dim=0))
+            losses['loss_bbox'] = loss_bbox
+
+            if self.with_corner_loss:
+                pos_roi_boxes3d = rois[..., 1:].view(-1, code_size)[pos_inds]
+                pos_roi_boxes3d = pos_roi_boxes3d.view(-1, code_size)
+                batch_anchors = pos_roi_boxes3d.clone().detach()
+                pos_rois_rotation = pos_roi_boxes3d[..., 6].view(-1)
+                roi_xyz = pos_roi_boxes3d[..., 0:3].view(-1, 3)
+                batch_anchors[..., 0:3] = 0
+                # decode boxes
+                pred_boxes3d = self.bbox_coder.decode(
+                    batch_anchors,
+                    pos_bbox_pred.view(-1, code_size)).view(-1, code_size)
+
+                pred_boxes3d[..., 0:3] = rotation_3d_in_axis(
+                    pred_boxes3d[..., 0:3].unsqueeze(1),
+                    pos_rois_rotation,
+                    axis=2).squeeze(1)
+
+                pred_boxes3d[:, 0:3] += roi_xyz
+
+                # calculate corner loss
+                loss_corner = self.get_corner_loss_lidar(
+                    pred_boxes3d, pos_gt_bboxes)
+                losses['loss_corner'] = loss_corner.mean()
+
+        return losses
+
+    def get_targets(self,
+                    sampling_results: SamplingResult,
+                    rcnn_train_cfg: dict,
+                    concat: bool = True) -> Tuple[torch.Tensor]:
+        """Generate targets.
+
+        Args:
+            sampling_results (list[:obj:`SamplingResult`]):
+                Sampled results from rois.
+            rcnn_train_cfg (:obj:`ConfigDict`): Training config of rcnn.
+            concat (bool): Whether to concatenate targets between batches.
+
+        Returns:
+            tuple[torch.Tensor]: Targets of boxes and class prediction.
+        """
+        pos_bboxes_list = [res.pos_bboxes for res in sampling_results]
+        pos_gt_bboxes_list = [res.pos_gt_bboxes for res in sampling_results]
+        iou_list = [res.iou for res in sampling_results]
+        targets = multi_apply(
+            self._get_target_single,
+            pos_bboxes_list,
+            pos_gt_bboxes_list,
+            iou_list,
+            cfg=rcnn_train_cfg)
+
+        (label, bbox_targets, pos_gt_bboxes, reg_mask, label_weights,
+         bbox_weights) = targets
+
+        if concat:
+            label = torch.cat(label, 0)
+            bbox_targets = torch.cat(bbox_targets, 0)
+            pos_gt_bboxes = torch.cat(pos_gt_bboxes, 0)
+            reg_mask = torch.cat(reg_mask, 0)
+
+            label_weights = torch.cat(label_weights, 0)
+            label_weights /= torch.clamp(label_weights.sum(), min=1.0)
+
+            bbox_weights = torch.cat(bbox_weights, 0)
+            bbox_weights /= torch.clamp(bbox_weights.sum(), min=1.0)
+
+        return (label, bbox_targets, pos_gt_bboxes, reg_mask, label_weights,
+                bbox_weights)
+
+    def _get_target_single(self, pos_bboxes: torch.Tensor,
+                           pos_gt_bboxes: torch.Tensor, ious: torch.Tensor,
+                           cfg: dict) -> Tuple[torch.Tensor]:
+        """Generate training targets for a single sample.
+
+        Args:
+            pos_bboxes (torch.Tensor): Positive boxes with shape
+                (N, 7).
+            pos_gt_bboxes (torch.Tensor): Ground truth boxes with shape
+                (M, 7).
+            ious (torch.Tensor): IoU between `pos_bboxes` and `pos_gt_bboxes`
+                in shape (N, M).
+            cfg (dict): Training configs.
+
+        Returns:
+            tuple[torch.Tensor]: Target for positive boxes.
+                (label, bbox_targets, pos_gt_bboxes, reg_mask, label_weights,
+                bbox_weights)
+        """
+        cls_pos_mask = ious > cfg.cls_pos_thr
+        cls_neg_mask = ious < cfg.cls_neg_thr
+        interval_mask = (cls_pos_mask == 0) & (cls_neg_mask == 0)
+
+        # iou regression target
+        label = (cls_pos_mask > 0).float()
+        label[interval_mask] = ious[interval_mask] * 2 - 0.5
+        # label weights
+        label_weights = (label >= 0).float()
+
+        # box regression target
+        reg_mask = pos_bboxes.new_zeros(ious.size(0)).long()
+        reg_mask[0:pos_gt_bboxes.size(0)] = 1
+        bbox_weights = (reg_mask > 0).float()
+        if reg_mask.bool().any():
+            pos_gt_bboxes_ct = pos_gt_bboxes.clone().detach()
+            roi_center = pos_bboxes[..., 0:3]
+            roi_ry = pos_bboxes[..., 6] % (2 * np.pi)
+
+            # canonical transformation
+            pos_gt_bboxes_ct[..., 0:3] -= roi_center
+            pos_gt_bboxes_ct[..., 6] -= roi_ry
+            pos_gt_bboxes_ct[..., 0:3] = rotation_3d_in_axis(
+                pos_gt_bboxes_ct[..., 0:3].unsqueeze(1), -roi_ry,
+                axis=2).squeeze(1)
+
+            # flip orientation if rois have opposite orientation
+            ry_label = pos_gt_bboxes_ct[..., 6] % (2 * np.pi)  # 0 ~ 2pi
+            opposite_flag = (ry_label > np.pi * 0.5) & (ry_label < np.pi * 1.5)
+            ry_label[opposite_flag] = (ry_label[opposite_flag] + np.pi) % (
+                2 * np.pi)  # (0 ~ pi/2, 3pi/2 ~ 2pi)
+            flag = ry_label > np.pi
+            ry_label[flag] = ry_label[flag] - np.pi * 2  # (-pi/2, pi/2)
+            ry_label = torch.clamp(ry_label, min=-np.pi / 2, max=np.pi / 2)
+            pos_gt_bboxes_ct[..., 6] = ry_label
+
+            rois_anchor = pos_bboxes.clone().detach()
+            rois_anchor[:, 0:3] = 0
+            rois_anchor[:, 6] = 0
+            bbox_targets = self.bbox_coder.encode(rois_anchor,
+                                                  pos_gt_bboxes_ct)
+        else:
+            # no fg bbox
+            bbox_targets = pos_gt_bboxes.new_empty((0, 7))
+
+        return (label, bbox_targets, pos_gt_bboxes, reg_mask, label_weights,
+                bbox_weights)
+
+    def get_corner_loss_lidar(self,
+                              pred_bbox3d: torch.Tensor,
+                              gt_bbox3d: torch.Tensor,
+                              delta: float = 1.0) -> torch.Tensor:
+        """Calculate corner loss of given boxes.
+
+        Args:
+            pred_bbox3d (torch.FloatTensor): Predicted boxes in shape (N, 7).
+            gt_bbox3d (torch.FloatTensor): Ground truth boxes in shape (N, 7).
+            delta (float, optional): huber loss threshold. Defaults to 1.0
+
+        Returns:
+            torch.FloatTensor: Calculated corner loss in shape (N).
+        """
+        assert pred_bbox3d.shape[0] == gt_bbox3d.shape[0]
+
+        # This is a little bit hack here because we assume the box for
+        # Part-A2 is in LiDAR coordinates
+        gt_boxes_structure = LiDARInstance3DBoxes(gt_bbox3d)
+        pred_box_corners = LiDARInstance3DBoxes(pred_bbox3d).corners
+        gt_box_corners = gt_boxes_structure.corners
+
+        # This flip only changes the heading direction of GT boxes
+        gt_bbox3d_flip = gt_boxes_structure.clone()
+        gt_bbox3d_flip.tensor[:, 6] += np.pi
+        gt_box_corners_flip = gt_bbox3d_flip.corners
+
+        corner_dist = torch.min(
+            torch.norm(pred_box_corners - gt_box_corners, dim=2),
+            torch.norm(pred_box_corners - gt_box_corners_flip,
+                       dim=2))  # (N, 8)
+        # huber loss
+        abs_error = torch.abs(corner_dist)
+        corner_loss = torch.where(abs_error < delta,
+                                  0.5 * abs_error**2 / delta,
+                                  abs_error - 0.5 * delta)
+        return corner_loss.mean(dim=1)
+
+    def get_results(self,
+                    rois: torch.Tensor,
+                    cls_preds: torch.Tensor,
+                    bbox_reg: torch.Tensor,
+                    class_labels: torch.Tensor,
+                    input_metas: List[dict],
+                    test_cfg: dict = None) -> InstanceList:
+        """Generate bboxes from bbox head predictions.
+
+        Args:
+            rois (torch.Tensor): Roi bounding boxes.
+            cls_preds (torch.Tensor): Scores of bounding boxes.
+            bbox_reg (torch.Tensor): Bounding boxes predictions
+            class_labels (torch.Tensor): Label of classes
+            input_metas (list[dict]): Point cloud meta info.
+            test_cfg (:obj:`ConfigDict`): Testing config.
+
+        Returns:
+            list[:obj:`InstanceData`]: Detection results of each sample
+            after the post process.
+            Each item usually contains following keys.
+
+            - scores_3d (Tensor): Classification scores, has a shape
+              (num_instances, )
+            - labels_3d (Tensor): Labels of bboxes, has a shape
+              (num_instances, ).
+            - bboxes_3d (BaseInstance3DBoxes): Prediction of bboxes,
+              contains a tensor with shape (num_instances, C), where
+              C >= 7.
+        """
+        roi_batch_id = rois[..., 0]
+        roi_boxes = rois[..., 1:]  # boxes without batch id
+        batch_size = int(roi_batch_id.max().item() + 1)
+
+        # decode boxes
+        roi_ry = roi_boxes[..., 6].view(-1)
+        roi_xyz = roi_boxes[..., 0:3].view(-1, 3)
+        local_roi_boxes = roi_boxes.clone().detach()
+        local_roi_boxes[..., 0:3] = 0
+        batch_box_preds = self.bbox_coder.decode(local_roi_boxes, bbox_reg)
+        batch_box_preds[..., 0:3] = rotation_3d_in_axis(
+            batch_box_preds[..., 0:3].unsqueeze(1), roi_ry, axis=2).squeeze(1)
+        batch_box_preds[:, 0:3] += roi_xyz
+
+        # post processing
+        result_list = []
+        for batch_id in range(batch_size):
+            cls_preds = cls_preds[roi_batch_id == batch_id]
+            box_preds = batch_box_preds[roi_batch_id == batch_id]
+            label_preds = class_labels[batch_id]
+
+            cls_preds = cls_preds.sigmoid()
+            cls_preds, _ = torch.max(cls_preds, dim=-1)
+            selected = self.class_agnostic_nms(
+                scores=cls_preds,
+                bbox_preds=box_preds,
+                input_meta=input_metas[batch_id],
+                nms_cfg=test_cfg)
+
+            selected_bboxes = box_preds[selected]
+            selected_label_preds = label_preds[selected]
+            selected_scores = cls_preds[selected]
+
+            results = InstanceData()
+            results.bboxes_3d = input_metas[batch_id]['box_type_3d'](
+                selected_bboxes, self.bbox_coder.code_size)
+            results.scores_3d = selected_scores
+            results.labels_3d = selected_label_preds
+
+            result_list.append(results)
+        return result_list
+
+    def class_agnostic_nms(self, scores: torch.Tensor,
+                           bbox_preds: torch.Tensor, nms_cfg: dict,
+                           input_meta: dict) -> Tuple[torch.Tensor]:
+        """Class agnostic NMS for box head.
+
+        Args:
+            scores (torch.Tensor): Object score of bounding boxes.
+            bbox_preds (torch.Tensor): Predicted bounding boxes.
+            nms_cfg (dict): NMS config dict.
+            input_meta (dict): Contain pcd and img's meta info.
+
+        Returns:
+            tuple[torch.Tensor]: Bounding boxes, scores and labels.
+        """
+        obj_scores = scores.clone()
+        if nms_cfg.use_rotate_nms:
+            nms_func = nms_bev
+        else:
+            nms_func = nms_normal_bev
+
+        bbox = input_meta['box_type_3d'](
+            bbox_preds.clone(),
+            box_dim=bbox_preds.shape[-1],
+            with_yaw=True,
+            origin=(0.5, 0.5, 0.5))
+
+        if nms_cfg.score_thr is not None:
+            scores_mask = (obj_scores >= nms_cfg.score_thr)
+            obj_scores = obj_scores[scores_mask]
+            bbox = bbox[scores_mask]
+        selected = []
+        if obj_scores.shape[0] > 0:
+            box_scores_nms, indices = torch.topk(
+                obj_scores, k=min(4096, obj_scores.shape[0]))
+            bbox_bev = bbox.bev[indices]
+            bbox_for_nms = xywhr2xyxyr(bbox_bev)
+
+            keep = nms_func(bbox_for_nms, box_scores_nms, nms_cfg.nms_thr)
+            selected = indices[keep]
+        if nms_cfg.score_thr is not None:
+            original_idxs = scores_mask.nonzero().view(-1)
+            selected = original_idxs[selected]
+        return selected

mmdet3d/models/roi_heads/mask_heads/__init__.py

 # Copyright (c) OpenMMLab. All rights reserved.
+from .foreground_segmentation_head import ForegroundSegmentationHead
 from .pointwise_semantic_head import PointwiseSemanticHead
 from .primitive_head import PrimitiveHead
 
-__all__ = ['PointwiseSemanticHead', 'PrimitiveHead']
+__all__ = [
+    'PointwiseSemanticHead', 'PrimitiveHead', 'ForegroundSegmentationHead'
+]

mmdet3d/models/roi_heads/mask_heads/foreground_segmentation_head.py

+# Copyright (c) OpenMMLab. All rights reserved.
+from typing import Dict, Optional, Tuple
+
+import torch
+from mmcv.cnn.bricks import build_norm_layer
+from mmengine.model import BaseModule
+from mmengine.structures import InstanceData
+from torch import nn as nn
+
+from mmdet3d.models.builder import build_loss
+from mmdet3d.registry import MODELS
+from mmdet3d.utils import InstanceList
+from mmdet.models.utils import multi_apply
+
+
+@MODELS.register_module()
+class ForegroundSegmentationHead(BaseModule):
+    """Foreground segmentation head.
+
+    Args:
+        in_channels (int): The number of input channel.
+        mlp_channels (tuple[int]): Specify of mlp channels. Defaults
+            to (256, 256).
+        extra_width (float): Boxes enlarge width. Default used 0.1.
+        norm_cfg (dict): Type of normalization method. Defaults to
+            dict(type='BN1d', eps=1e-5, momentum=0.1).
+        init_cfg (dict, optional): Initialize config of
+            model. Defaults to None.
+        loss_seg (dict): Config of segmentation loss. Defaults to
+            dict(type='mmdet.FocalLoss')
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        mlp_channels: Tuple[int] = (256, 256),
+        extra_width: float = 0.1,
+        norm_cfg: dict = dict(type='BN1d', eps=1e-5, momentum=0.1),
+        init_cfg: Optional[dict] = None,
+        loss_seg: dict = dict(
+            type='mmdet.FocalLoss',
+            use_sigmoid=True,
+            reduction='sum',
+            gamma=2.0,
+            alpha=0.25,
+            activated=True,
+            loss_weight=1.0)
+    ) -> None:
+        super(ForegroundSegmentationHead, self).__init__(init_cfg=init_cfg)
+        self.extra_width = extra_width
+        self.num_classes = 1
+
+        self.in_channels = in_channels
+        self.use_sigmoid_cls = loss_seg.get('use_sigmoid', False)
+
+        out_channels = 1
+        if self.use_sigmoid_cls:
+            self.out_channels = out_channels
+        else:
+            self.out_channels = out_channels + 1
+
+        mlps_layers = []
+        cin = in_channels
+        for mlp in mlp_channels:
+            mlps_layers.extend([
+                nn.Linear(cin, mlp, bias=False),
+                build_norm_layer(norm_cfg, mlp)[1],
+                nn.ReLU()
+            ])
+            cin = mlp
+        mlps_layers.append(nn.Linear(cin, self.out_channels, bias=True))
+
+        self.seg_cls_layer = nn.Sequential(*mlps_layers)
+
+        self.loss_seg = build_loss(loss_seg)
+
+    def forward(self, feats: torch.Tensor) -> dict:
+        """Forward head.
+
+        Args:
+            feats (torch.Tensor): Point-wise features.
+
+        Returns:
+            dict: Segment predictions.
+        """
+        seg_preds = self.seg_cls_layer(feats)
+        return dict(seg_preds=seg_preds)
+
+    def _get_targets_single(self, point_xyz: torch.Tensor,
+                            gt_bboxes_3d: InstanceData,
+                            gt_labels_3d: torch.Tensor) -> torch.Tensor:
+        """generate segmentation targets for a single sample.
+
+        Args:
+            point_xyz (torch.Tensor): Coordinate of points.
+            gt_bboxes_3d (:obj:`BaseInstance3DBoxes`): Ground truth boxes in
+                shape (box_num, 7).
+            gt_labels_3d (torch.Tensor): Class labels of ground truths in
+                shape (box_num).
+
+        Returns:
+            torch.Tensor: Points class labels.
+        """
+        point_cls_labels_single = point_xyz.new_zeros(
+            point_xyz.shape[0]).long()
+        enlarged_gt_boxes = gt_bboxes_3d.enlarged_box(self.extra_width)
+
+        box_idxs_of_pts = gt_bboxes_3d.points_in_boxes_part(point_xyz).long()
+        extend_box_idxs_of_pts = enlarged_gt_boxes.points_in_boxes_part(
+            point_xyz).long()
+        box_fg_flag = box_idxs_of_pts >= 0
+        fg_flag = box_fg_flag.clone()
+        ignore_flag = fg_flag ^ (extend_box_idxs_of_pts >= 0)
+        point_cls_labels_single[ignore_flag] = -1
+        gt_box_of_fg_points = gt_labels_3d[box_idxs_of_pts[fg_flag]]
+        point_cls_labels_single[
+            fg_flag] = 1 if self.num_classes == 1 else\
+            gt_box_of_fg_points.long()
+        return point_cls_labels_single,
+
+    def get_targets(self, points_bxyz: torch.Tensor,
+                    batch_gt_instances_3d: InstanceList) -> dict:
+        """Generate segmentation targets.
+
+        Args:
+            points_bxyz (torch.Tensor): The coordinates of point in shape
+                (B, num_points, 3).
+            batch_gt_instances_3d (list[:obj:`InstanceData`]): Batch of
+                gt_instances. It usually includes ``bboxes_3d`` and
+                ``labels_3d`` attributes.
+
+        Returns:
+            dict: Prediction targets
+                - seg_targets (torch.Tensor): Segmentation targets.
+        """
+        batch_size = len(batch_gt_instances_3d)
+        points_xyz_list = []
+        gt_bboxes_3d = []
+        gt_labels_3d = []
+        for idx in range(batch_size):
+            coords_idx = points_bxyz[:, 0] == idx
+            points_xyz_list.append(points_bxyz[coords_idx][..., 1:])
+            gt_bboxes_3d.append(batch_gt_instances_3d[idx].bboxes_3d)
+            gt_labels_3d.append(batch_gt_instances_3d[idx].labels_3d)
+        seg_targets, = multi_apply(self._get_targets_single, points_xyz_list,
+                                   gt_bboxes_3d, gt_labels_3d)
+        seg_targets = torch.cat(seg_targets, dim=0)
+        return dict(seg_targets=seg_targets)
+
+    def loss(self, semantic_results: dict,
+             semantic_targets: dict) -> Dict[str, torch.Tensor]:
+        """Calculate point-wise segmentation losses.
+
+        Args:
+            semantic_results (dict): Results from semantic head.
+            semantic_targets (dict): Targets of semantic results.
+
+        Returns:
+            dict: Loss of segmentation.
+
+            - loss_semantic (torch.Tensor): Segmentation prediction loss.
+        """
+        seg_preds = semantic_results['seg_preds']
+        seg_targets = semantic_targets['seg_targets']
+
+        positives = (seg_targets > 0)
+
+        negative_cls_weights = (seg_targets == 0).float()
+        seg_weights = (negative_cls_weights + 1.0 * positives).float()
+        pos_normalizer = positives.sum(dim=0).float()
+        seg_weights /= torch.clamp(pos_normalizer, min=1.0)
+
+        seg_preds = torch.sigmoid(seg_preds)
+        loss_seg = self.loss_seg(seg_preds, (~positives).long(), seg_weights)
+        return dict(loss_semantic=loss_seg)

mmdet3d/models/roi_heads/pv_rcnn_roi_head.py

+# Copyright (c) OpenMMLab. All rights reserved.
+from typing import List, Optional
+
+import torch
+from torch.nn import functional as F
+
+from mmdet3d.models.roi_heads.base_3droi_head import Base3DRoIHead
+from mmdet3d.registry import MODELS
+from mmdet3d.structures import bbox3d2roi
+from mmdet3d.structures.det3d_data_sample import SampleList
+from mmdet3d.utils import InstanceList
+from mmdet.models.task_modules import AssignResult
+from mmdet.models.task_modules.samplers import SamplingResult
+
+
+@MODELS.register_module()
+class PVRCNNRoiHead(Base3DRoIHead):
+    """RoI head for PV-RCNN.
+
+    Args:
+        num_classes (int): The number of classes. Defaults to 3.
+        semantic_head (dict, optional): Config of semantic head.
+            Defaults to None.
+        bbox_roi_extractor (dict, optional): Config of roi_extractor.
+            Defaults to None.
+        bbox_head (dict, optional): Config of bbox_head. Defaults to None.
+        train_cfg (dict, optional): Train config of model.
+            Defaults to None.
+        test_cfg (dict, optional): Train config of model.
+            Defaults to None.
+        init_cfg (dict, optional): Initialize config of
+            model. Defaults to None.
+    """
+
+    def __init__(self,
+                 num_classes: int = 3,
+                 semantic_head: Optional[dict] = None,
+                 bbox_roi_extractor: Optional[dict] = None,
+                 bbox_head: Optional[dict] = None,
+                 train_cfg: Optional[dict] = None,
+                 test_cfg: Optional[dict] = None,
+                 init_cfg: Optional[dict] = None):
+        super(PVRCNNRoiHead, self).__init__(
+            bbox_head=bbox_head,
+            bbox_roi_extractor=bbox_roi_extractor,
+            train_cfg=train_cfg,
+            test_cfg=test_cfg,
+            init_cfg=init_cfg)
+        self.num_classes = num_classes
+        self.semantic_head = MODELS.build(semantic_head)
+
+        self.init_assigner_sampler()
+
+    @property
+    def with_semantic(self):
+        """bool: whether the head has semantic branch"""
+        return hasattr(self,
+                       'semantic_head') and self.semantic_head is not None
+
+    def loss(self, feats_dict: dict, rpn_results_list: InstanceList,
+             batch_data_samples: SampleList, **kwargs) -> dict:
+        """Training forward function of PVRCNNROIHead.
+
+        Args:
+            feats_dict (dict): Contains point-wise features.
+            rpn_results_list (List[:obj:`InstanceData`]): Detection results
+                of rpn head.
+            batch_data_samples (List[:obj:`Det3DDataSample`]): The Data
+                samples. It usually includes information such as
+                `gt_instance_3d`, `gt_panoptic_seg_3d` and `gt_sem_seg_3d`.
+
+        Returns:
+            dict: losses from each head.
+
+            - loss_semantic (torch.Tensor): loss of semantic head.
+            - loss_bbox (torch.Tensor): loss of bboxes.
+            - loss_cls (torch.Tensor): loss of object classification.
+            - loss_corner (torch.Tensor): loss of bboxes corners.
+        """
+        losses = dict()
+        batch_gt_instances_3d = []
+        batch_gt_instances_ignore = []
+        for data_sample in batch_data_samples:
+            batch_gt_instances_3d.append(data_sample.gt_instances_3d)
+            if 'ignored_instances' in data_sample:
+                batch_gt_instances_ignore.append(data_sample.ignored_instances)
+            else:
+                batch_gt_instances_ignore.append(None)
+        if self.with_semantic:
+            semantic_results = self._semantic_forward_train(
+                feats_dict['keypoint_features'], feats_dict['keypoints'],
+                batch_gt_instances_3d)
+            losses['loss_semantic'] = semantic_results['loss_semantic']
+
+        sample_results = self._assign_and_sample(rpn_results_list,
+                                                 batch_gt_instances_3d)
+        if self.with_bbox:
+            bbox_results = self._bbox_forward_train(
+                semantic_results['seg_preds'],
+                feats_dict['fusion_keypoint_features'],
+                feats_dict['keypoints'], sample_results)
+            losses.update(bbox_results['loss_bbox'])
+
+        return losses
+
+    def predict(self, feats_dict: dict, rpn_results_list: InstanceList,
+                batch_data_samples: SampleList, **kwargs) -> SampleList:
+        """Perform forward propagation of the roi head and predict detection
+        results on the features of the upstream network.
+
+        Args:
+            feats_dict (dict): Contains point-wise features.
+            rpn_results_list (List[:obj:`InstanceData`]): Detection results
+                of rpn head.
+            batch_data_samples (List[:obj:`Det3DDataSample`]): The Data
+                samples. It usually includes information such as
+                `gt_instance_3d`, `gt_panoptic_seg_3d` and `gt_sem_seg_3d`.
+
+        Returns:
+            list[:obj:`InstanceData`]: Detection results of each sample
+            after the post process.
+            Each item usually contains following keys.
+
+            - scores_3d (Tensor): Classification scores, has a shape
+              (num_instances, )
+            - labels_3d (Tensor): Labels of bboxes, has a shape
+              (num_instances, ).
+            - bboxes_3d (BaseInstance3DBoxes): Prediction of bboxes,
+              contains a tensor with shape (num_instances, C), where
+              C >= 7.
+        """
+        assert self.with_bbox, 'Bbox head must be implemented.'
+        assert self.with_semantic, 'Semantic head must be implemented.'
+
+        batch_input_metas = [
+            data_samples.metainfo for data_samples in batch_data_samples
+        ]
+
+        semantic_results = self.semantic_head(feats_dict['keypoint_features'])
+        point_features = feats_dict[
+            'fusion_keypoint_features'] * semantic_results[
+                'seg_preds'].sigmoid().max(
+                    dim=-1, keepdim=True).values
+        rois = bbox3d2roi(
+            [res['bboxes_3d'].tensor for res in rpn_results_list])
+        labels_3d = [res['labels_3d'] for res in rpn_results_list]
+        bbox_results = self._bbox_forward(point_features,
+                                          feats_dict['keypoints'], rois)
+
+        results_list = self.bbox_head.get_results(rois,
+                                                  bbox_results['bbox_scores'],
+                                                  bbox_results['bbox_reg'],
+                                                  labels_3d, batch_input_metas,
+                                                  self.test_cfg)
+        return results_list
+
+    def _bbox_forward_train(self, seg_preds: torch.Tensor,
+                            keypoint_features: torch.Tensor,
+                            keypoints: torch.Tensor,
+                            sampling_results: SamplingResult) -> dict:
+        """Forward training function of roi_extractor and bbox_head.
+
+        Args:
+            seg_preds (torch.Tensor): Point-wise semantic features.
+            keypoint_features (torch.Tensor): key points features
+                from points encoder.
+            keypoints (torch.Tensor): Coordinate of key points.
+            sampling_results (:obj:`SamplingResult`): Sampled results used
+                for training.
+
+        Returns:
+            dict: Forward results including losses and predictions.
+        """
+        rois = bbox3d2roi([res.bboxes for res in sampling_results])
+        keypoint_features = keypoint_features * seg_preds.sigmoid().max(
+            dim=-1, keepdim=True).values
+        bbox_results = self._bbox_forward(keypoint_features, keypoints, rois)
+
+        bbox_targets = self.bbox_head.get_targets(sampling_results,
+                                                  self.train_cfg)
+        loss_bbox = self.bbox_head.loss(bbox_results['bbox_scores'],
+                                        bbox_results['bbox_reg'], rois,
+                                        *bbox_targets)
+
+        bbox_results.update(loss_bbox=loss_bbox)
+        return bbox_results
+
+    def _bbox_forward(self, keypoint_features: torch.Tensor,
+                      keypoints: torch.Tensor, rois: torch.Tensor) -> dict:
+        """Forward function of roi_extractor and bbox_head used in both
+        training and testing.
+
+        Args:
+            rois (Tensor): Roi boxes.
+            keypoint_features (torch.Tensor): key points features
+                from points encoder.
+            keypoints (torch.Tensor): Coordinate of key points.
+            rois (Tensor): Roi boxes.
+
+        Returns:
+            dict: Contains predictions of bbox_head and
+                features of roi_extractor.
+        """
+        pooled_keypoint_features = self.bbox_roi_extractor(
+            keypoint_features, keypoints[..., 1:], keypoints[..., 0].int(),
+            rois)
+        bbox_score, bbox_reg = self.bbox_head(pooled_keypoint_features)
+
+        bbox_results = dict(bbox_scores=bbox_score, bbox_reg=bbox_reg)
+        return bbox_results
+
+    def _assign_and_sample(
+            self, proposal_list: InstanceList,
+            batch_gt_instances_3d: InstanceList) -> List[SamplingResult]:
+        """Assign and sample proposals for training.
+
+        Args:
+            proposal_list (list[:obj:`InstancesData`]): Proposals produced by
+                rpn head.
+            batch_gt_instances_3d (list[:obj:`InstanceData`]): Batch of
+                gt_instances. It usually includes ``bboxes_3d`` and
+                ``labels_3d`` attributes.
+
+        Returns:
+            list[:obj:`SamplingResult`]: Sampled results of each training
+                sample.
+        """
+        sampling_results = []
+        # bbox assign
+        for batch_idx in range(len(proposal_list)):
+            cur_proposal_list = proposal_list[batch_idx]
+            cur_boxes = cur_proposal_list['bboxes_3d']
+            cur_labels_3d = cur_proposal_list['labels_3d']
+            cur_gt_instances_3d = batch_gt_instances_3d[batch_idx]
+            cur_gt_instances_3d.bboxes_3d = cur_gt_instances_3d.\
+                bboxes_3d.tensor
+            cur_gt_bboxes = batch_gt_instances_3d[batch_idx].bboxes_3d.to(
+                cur_boxes.device)
+            cur_gt_labels = batch_gt_instances_3d[batch_idx].labels_3d
+
+            batch_num_gts = 0
+            # 0 is bg
+            batch_gt_indis = cur_gt_labels.new_full((len(cur_boxes), ), 0)
+            batch_max_overlaps = cur_boxes.tensor.new_zeros(len(cur_boxes))
+            # -1 is bg
+            batch_gt_labels = cur_gt_labels.new_full((len(cur_boxes), ), -1)
+
+            # each class may have its own assigner
+            if isinstance(self.bbox_assigner, list):
+                for i, assigner in enumerate(self.bbox_assigner):
+                    gt_per_cls = (cur_gt_labels == i)
+                    pred_per_cls = (cur_labels_3d == i)
+                    cur_assign_res = assigner.assign(
+                        cur_proposal_list[pred_per_cls],
+                        cur_gt_instances_3d[gt_per_cls])
+                    # gather assign_results in different class into one result
+                    batch_num_gts += cur_assign_res.num_gts
+                    # gt inds (1-based)
+                    gt_inds_arange_pad = gt_per_cls.nonzero(
+                        as_tuple=False).view(-1) + 1
+                    # pad 0 for indice unassigned
+                    gt_inds_arange_pad = F.pad(
+                        gt_inds_arange_pad, (1, 0), mode='constant', value=0)
+                    # pad -1 for indice ignore
+                    gt_inds_arange_pad = F.pad(
+                        gt_inds_arange_pad, (1, 0), mode='constant', value=-1)
+                    # convert to 0~gt_num+2 for indices
+                    gt_inds_arange_pad += 1
+                    # now 0 is bg, >1 is fg in batch_gt_indis
+                    batch_gt_indis[pred_per_cls] = gt_inds_arange_pad[
+                        cur_assign_res.gt_inds + 1] - 1
+                    batch_max_overlaps[
+                        pred_per_cls] = cur_assign_res.max_overlaps
+                    batch_gt_labels[pred_per_cls] = cur_assign_res.labels
+
+                assign_result = AssignResult(batch_num_gts, batch_gt_indis,
+                                             batch_max_overlaps,
+                                             batch_gt_labels)
+            else:  # for single class
+                assign_result = self.bbox_assigner.assign(
+                    cur_proposal_list, cur_gt_instances_3d)
+            # sample boxes
+            sampling_result = self.bbox_sampler.sample(assign_result,
+                                                       cur_boxes.tensor,
+                                                       cur_gt_bboxes,
+                                                       cur_gt_labels)
+            sampling_results.append(sampling_result)
+        return sampling_results
+
+    def _semantic_forward_train(self, keypoint_features: torch.Tensor,
+                                keypoints: torch.Tensor,
+                                batch_gt_instances_3d: InstanceList) -> dict:
+        """Train semantic head.
+
+        Args:
+            keypoint_features (torch.Tensor): key points features
+                from points encoder.
+            keypoints (torch.Tensor): Coordinate of key points.
+            batch_gt_instances_3d (list[:obj:`InstanceData`]): Batch of
+                gt_instances. It usually includes ``bboxes_3d`` and
+                ``labels_3d`` attributes.
+
+        Returns:
+            dict: Segmentation results including losses
+        """
+        semantic_results = self.semantic_head(keypoint_features)
+        semantic_targets = self.semantic_head.get_targets(
+            keypoints, batch_gt_instances_3d)
+        loss_semantic = self.semantic_head.loss(semantic_results,
+                                                semantic_targets)
+        semantic_results.update(loss_semantic)
+        return semantic_results

mmdet3d/models/roi_heads/roi_extractors/__init__.py

 # Copyright (c) OpenMMLab. All rights reserved.
 from mmdet.models.roi_heads.roi_extractors import SingleRoIExtractor
 
+from .batch_roigridpoint_extractor import Batch3DRoIGridExtractor
 from .single_roiaware_extractor import Single3DRoIAwareExtractor
 from .single_roipoint_extractor import Single3DRoIPointExtractor
 
 __all__ = [
     'SingleRoIExtractor', 'Single3DRoIAwareExtractor',
-    'Single3DRoIPointExtractor'
+    'Single3DRoIPointExtractor', 'Batch3DRoIGridExtractor'
 ]

mmdet3d/models/roi_heads/roi_extractors/batch_roigridpoint_extractor.py

+# Copyright (c) OpenMMLab. All rights reserved.
+import torch
+from mmengine.model import BaseModule
+
+from mmdet3d.registry import MODELS
+from mmdet3d.structures.bbox_3d import rotation_3d_in_axis
+
+
+@MODELS.register_module()
+class Batch3DRoIGridExtractor(BaseModule):
+    """Grid point wise roi-aware Extractor.
+
+    Args:
+        grid_size (int): The number of grid points in a roi bbox.
+            Defaults to 6.
+        roi_layer (dict, optional): Config of sa module to get
+            grid points features. Defaults to None.
+        init_cfg (dict, optional): Initialize config of
+            model. Defaults to None.
+    """
+
+    def __init__(self,
+                 grid_size: int = 6,
+                 roi_layer: dict = None,
+                 init_cfg: dict = None) -> None:
+        super(Batch3DRoIGridExtractor, self).__init__(init_cfg=init_cfg)
+        self.roi_grid_pool_layer = MODELS.build(roi_layer)
+        self.grid_size = grid_size
+
+    def forward(self, feats: torch.Tensor, coordinate: torch.Tensor,
+                batch_inds: torch.Tensor, rois: torch.Tensor) -> torch.Tensor:
+        """Forward roi extractor to extract grid points feature.
+
+        Args:
+            feats (torch.Tensor): Key points features.
+            coordinate (torch.Tensor): Key points coordinates.
+            batch_inds (torch.Tensor): Input batch indexes.
+            rois (torch.Tensor): Detection results of rpn head.
+
+        Returns:
+            torch.Tensor: Grid points features.
+        """
+        batch_size = int(batch_inds.max()) + 1
+
+        xyz = coordinate
+        xyz_batch_cnt = xyz.new_zeros(batch_size).int()
+        for k in range(batch_size):
+            xyz_batch_cnt[k] = (batch_inds == k).sum()
+
+        rois_batch_inds = rois[:, 0].int()
+        # (N1+N2+..., 6x6x6, 3)
+        roi_grid = self.get_dense_grid_points(rois[:, 1:])
+
+        new_xyz = roi_grid.view(-1, 3)
+        new_xyz_batch_cnt = new_xyz.new_zeros(batch_size).int()
+        for k in range(batch_size):
+            new_xyz_batch_cnt[k] = ((rois_batch_inds == k).sum() *
+                                    roi_grid.size(1))
+        pooled_points, pooled_features = self.roi_grid_pool_layer(
+            xyz=xyz.contiguous(),
+            xyz_batch_cnt=xyz_batch_cnt,
+            new_xyz=new_xyz.contiguous(),
+            new_xyz_batch_cnt=new_xyz_batch_cnt,
+            features=feats.contiguous())  # (M1 + M2 ..., C)
+
+        pooled_features = pooled_features.view(-1, self.grid_size,
+                                               self.grid_size, self.grid_size,
+                                               pooled_features.shape[-1])
+        # (BxN, 6, 6, 6, C)
+        return pooled_features
+
+    def get_dense_grid_points(self, rois: torch.Tensor) -> torch.Tensor:
+        """Get dense grid points from rois.
+
+        Args:
+            rois (torch.Tensor): Detection results of rpn head.
+
+        Returns:
+            torch.Tensor: Grid points coordinates.
+        """
+        rois_bbox = rois.clone()
+        rois_bbox[:, 2] += rois_bbox[:, 5] / 2
+        faked_features = rois_bbox.new_ones(
+            (self.grid_size, self.grid_size, self.grid_size))
+        dense_idx = faked_features.nonzero()
+        dense_idx = dense_idx.repeat(rois_bbox.size(0), 1, 1).float()
+        dense_idx = ((dense_idx + 0.5) / self.grid_size)
+        dense_idx[..., :3] -= 0.5
+
+        roi_ctr = rois_bbox[:, :3]
+        roi_dim = rois_bbox[:, 3:6]
+        roi_grid_points = dense_idx * roi_dim.view(-1, 1, 3)
+        roi_grid_points = rotation_3d_in_axis(
+            roi_grid_points, rois_bbox[:, 6], axis=2)
+        roi_grid_points += roi_ctr.view(-1, 1, 3)
+
+        return roi_grid_points

tests/test_models/test_detectors/test_pvrcnn.py

+import unittest
+
+import torch
+from mmengine import DefaultScope
+
+from mmdet3d.registry import MODELS
+from tests.utils.model_utils import (_create_detector_inputs,
+                                     _get_detector_cfg, _setup_seed)
+
+
+class TestPVRCNN(unittest.TestCase):
+
+    def test_pvrcnn(self):
+        import mmdet3d.models
+
+        assert hasattr(mmdet3d.models, 'PointVoxelRCNN')
+        DefaultScope.get_instance('test_pvrcnn', scope_name='mmdet3d')
+        _setup_seed(0)
+        pvrcnn_cfg = _get_detector_cfg(
+            'pvrcnn/pvrcnn_8xb2-80e_kitti-3d-3class.py')
+        model = MODELS.build(pvrcnn_cfg)
+        num_gt_instance = 2
+        packed_inputs = _create_detector_inputs(
+            num_gt_instance=num_gt_instance)
+
+        # TODO: Support aug data test
+        # aug_packed_inputs = [
+        #     _create_detector_inputs(num_gt_instance=num_gt_instance),
+        #     _create_detector_inputs(num_gt_instance=num_gt_instance + 1)
+        # ]
+        # test_aug_test
+        # metainfo = {
+        #     'pcd_scale_factor': 1,
+        #     'pcd_horizontal_flip': 1,
+        #     'pcd_vertical_flip': 1,
+        #     'box_type_3d': LiDARInstance3DBoxes
+        # }
+        # for item in aug_packed_inputs:
+        #     for batch_id in len(item['data_samples']):
+        #         item['data_samples'][batch_id].set_metainfo(metainfo)
+
+        if torch.cuda.is_available():
+            model = model.cuda()
+            # test simple_test
+            with torch.no_grad():
+                data = model.data_preprocessor(packed_inputs, True)
+                torch.cuda.empty_cache()
+                results = model.forward(**data, mode='predict')
+            self.assertEqual(len(results), 1)
+            self.assertIn('bboxes_3d', results[0].pred_instances_3d)
+            self.assertIn('scores_3d', results[0].pred_instances_3d)
+            self.assertIn('labels_3d', results[0].pred_instances_3d)
+
+            # save the memory
+            with torch.no_grad():
+                losses = model.forward(**data, mode='loss')
+                torch.cuda.empty_cache()
+            self.assertGreater(losses['loss_rpn_cls'][0], 0)
+            self.assertGreaterEqual(losses['loss_rpn_bbox'][0], 0)
+            self.assertGreaterEqual(losses['loss_rpn_dir'][0], 0)
+            self.assertGreater(losses['loss_semantic'], 0)
+            self.assertGreaterEqual(losses['loss_bbox'], 0)
+            self.assertGreaterEqual(losses['loss_cls'], 0)
+            self.assertGreaterEqual(losses['loss_corner'], 0)