Merge branch 'feature_semantic_head' into 'master'

semantic head with unittest

See merge request open-mmlab/mmdet.3d!13

mmdet3d/core/bbox/box_torch_ops.py

@@ -190,3 +190,23 @@ def rotation_2d(points, angles):
     rot_cos = torch.cos(angles)
     rot_mat_T = torch.stack([[rot_cos, -rot_sin], [rot_sin, rot_cos]])
     return torch.einsum('aij,jka->aik', points, rot_mat_T)
+
+
+def enlarge_box3d_lidar(boxes3d, extra_width):
+    """Enlarge the length, width and height of input boxes
+
+    Args:
+        boxes3d (torch.float32 or numpy.float32): bottom_center with
+            shape [N, 7], (x, y, z, w, l, h, ry) in LiDAR coords
+        extra_width (float): a fix number to add
+
+    Returns:
+        torch.float32 or numpy.float32: enlarged boxes
+    """
+    if isinstance(boxes3d, np.ndarray):
+        large_boxes3d = boxes3d.copy()
+    else:
+        large_boxes3d = boxes3d.clone()
+    large_boxes3d[:, 3:6] += extra_width * 2
+    large_boxes3d[:, 2] -= extra_width  # bottom center z minus extra_width
+    return large_boxes3d

mmdet3d/models/losses/__init__.py

-from mmdet.models.losses import FocalLoss, SmoothL1Loss
+from mmdet.models.losses import FocalLoss, SmoothL1Loss, binary_cross_entropy
 
-__all__ = ['FocalLoss', 'SmoothL1Loss']
+__all__ = ['FocalLoss', 'SmoothL1Loss', 'binary_cross_entropy']

mmdet3d/models/roi_heads/__init__.py

+from .mask_heads import PointwiseSemanticHead
+
+__all__ = ['PointwiseSemanticHead']

mmdet3d/models/roi_heads/mask_heads/__init__.py

+from .pointwise_semantic_head import PointwiseSemanticHead
+
+__all__ = ['PointwiseSemanticHead']

mmdet3d/models/roi_heads/mask_heads/pointwise_semantic_head.py

+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from mmdet3d.core import multi_apply
+from mmdet3d.core.bbox import box_torch_ops
+from mmdet3d.models.builder import build_loss
+from mmdet3d.ops.roiaware_pool3d import points_in_boxes_gpu
+from mmdet.models import HEADS
+
+
+@HEADS.register_module
+class PointwiseSemanticHead(nn.Module):
+    """Semantic segmentation head for point-wise segmentation.
+
+    Predict point-wise segmentation and part regression results for PartA2.
+    See https://arxiv.org/abs/1907.03670 for more detials.
+
+    Args:
+        in_channels (int): the number of input channel.
+        num_classes (int): the number of class.
+        extra_width (float): boxes enlarge width.
+        loss_seg (dict): Config of segmentation loss.
+        loss_part (dict): Config of part prediction loss.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 num_classes=3,
+                 extra_width=0.2,
+                 seg_score_thr=0.3,
+                 loss_seg=dict(
+                     type='FocalLoss',
+                     use_sigmoid=True,
+                     reduction='sum',
+                     gamma=2.0,
+                     alpha=0.25,
+                     loss_weight=1.0),
+                 loss_part=dict(
+                     type='CrossEntropyLoss',
+                     use_sigmoid=True,
+                     loss_weight=1.0)):
+        super(PointwiseSemanticHead, self).__init__()
+        self.extra_width = extra_width
+        self.num_classes = num_classes
+        self.seg_score_thr = seg_score_thr
+        self.seg_cls_layer = nn.Linear(in_channels, 1, bias=True)
+        self.seg_reg_layer = nn.Linear(in_channels, 3, bias=True)
+
+        self.loss_seg = build_loss(loss_seg)
+        self.loss_part = build_loss(loss_part)
+
+    def forward(self, x):
+        seg_preds = self.seg_cls_layer(x)  # (N, 1)
+        part_preds = self.seg_reg_layer(x)  # (N, 3)
+
+        seg_scores = torch.sigmoid(seg_preds).detach()
+        seg_mask = (seg_scores > self.seg_score_thr)
+
+        part_offsets = torch.sigmoid(part_preds).clone().detach()
+        part_offsets[seg_mask.view(-1) == 0] = 0
+        part_feats = torch.cat((part_offsets, seg_scores),
+                               dim=-1)  # shape (npoints, 4)
+        return dict(
+            seg_preds=seg_preds, part_preds=part_preds, part_feats=part_feats)
+
+    def get_targets_single(self, voxel_centers, gt_bboxes_3d, gt_labels_3d):
+        """generate segmentation and part prediction targets
+
+        Args:
+            voxel_centers (torch.Tensor): shape [voxel_num, 3],
+                the center of voxels
+            gt_bboxes_3d (torch.Tensor): shape [box_num, 7], gt boxes
+            gt_labels_3d (torch.Tensor): shape [box_num], class label of gt
+
+        Returns:
+            tuple : segmentation targets with shape [voxel_num]
+                part prediction targets with shape [voxel_num, 3]
+        """
+        enlarged_gt_boxes = box_torch_ops.enlarge_box3d_lidar(
+            gt_bboxes_3d, extra_width=self.extra_width)
+        part_targets = voxel_centers.new_zeros((voxel_centers.shape[0], 3),
+                                               dtype=torch.float32)
+        box_idx = points_in_boxes_gpu(
+            voxel_centers.unsqueeze(0),
+            gt_bboxes_3d.unsqueeze(0)).squeeze(0)  # -1 ~ box_num
+        enlarge_box_idx = points_in_boxes_gpu(
+            voxel_centers.unsqueeze(0),
+            enlarged_gt_boxes.unsqueeze(0)).squeeze(0).long()  # -1 ~ box_num
+
+        gt_labels_pad = F.pad(
+            gt_labels_3d, (1, 0), mode='constant', value=self.num_classes)
+        seg_targets = gt_labels_pad[(box_idx.long() + 1)]
+        fg_pt_flag = box_idx > -1
+        ignore_flag = fg_pt_flag ^ (enlarge_box_idx > -1)
+        seg_targets[ignore_flag] = -1
+
+        for k in range(gt_bboxes_3d.shape[0]):
+            k_box_flag = box_idx == k
+            # no point in current box (caused by velodyne reduce)
+            if not k_box_flag.any():
+                continue
+            fg_voxels = voxel_centers[k_box_flag]
+            transformed_voxels = fg_voxels - gt_bboxes_3d[k, 0:3]
+            transformed_voxels = box_torch_ops.rotation_3d_in_axis(
+                transformed_voxels.unsqueeze(0),
+                -gt_bboxes_3d[k, 6].view(1),
+                axis=2)
+            part_targets[k_box_flag] = transformed_voxels / gt_bboxes_3d[
+                k, 3:6] + voxel_centers.new_tensor([0.5, 0.5, 0])
+
+        part_targets = torch.clamp(part_targets, min=0)
+        return seg_targets, part_targets
+
+    def get_targets(self, voxels_dict, gt_bboxes_3d, gt_labels_3d):
+        batch_size = len(gt_labels_3d)
+        voxel_center_list = []
+        for idx in range(batch_size):
+            coords_idx = voxels_dict['coors'][:, 0] == idx
+            voxel_center_list.append(voxels_dict['voxel_centers'][coords_idx])
+
+        seg_targets, part_targets = multi_apply(self.get_targets_single,
+                                                voxel_center_list,
+                                                gt_bboxes_3d, gt_labels_3d)
+        seg_targets = torch.cat(seg_targets, dim=0)
+        part_targets = torch.cat(part_targets, dim=0)
+        return dict(seg_targets=seg_targets, part_targets=part_targets)
+
+    def loss(self, seg_preds, part_preds, seg_targets, part_targets):
+        """Calculate point-wise segmentation and part prediction losses.
+
+        Args:
+            seg_preds (torch.Tensor): prediction of binary
+                segmentation with shape [voxel_num, 1].
+            part_preds (torch.Tensor): prediction of part
+                with shape [voxel_num, 3].
+            seg_targets (torch.Tensor): target of segmentation
+                with shape [voxel_num, 1].
+            part_targets (torch.Tensor): target of part with
+                shape [voxel_num, 3].
+
+        Returns:
+            dict: loss of segmentation and part prediction.
+        """
+        pos_mask = (seg_targets > -1) & (seg_targets < self.num_classes)
+        binary_seg_target = pos_mask.long()
+        pos = pos_mask.float()
+        neg = (seg_targets == self.num_classes).float()
+        seg_weights = pos + neg
+        pos_normalizer = pos.sum()
+        seg_weights = seg_weights / torch.clamp(pos_normalizer, min=1.0)
+        loss_seg = self.loss_seg(seg_preds, binary_seg_target, seg_weights)
+
+        if pos_normalizer > 0:
+            loss_part = self.loss_part(part_preds[pos_mask],
+                                       part_targets[pos_mask])
+        else:
+            # fake a part loss
+            loss_part = loss_seg.new_tensor(0)
+
+        return dict(loss_seg=loss_seg, loss_part=loss_part)

tests/test_roiaware_pool3d.py

@@ -6,8 +6,8 @@ from mmdet3d.ops.roiaware_pool3d import (RoIAwarePool3d, points_in_boxes_cpu,
 
 
 def test_RoIAwarePool3d():
-    if not torch.cuda.is_available(
-    ):  # RoIAwarePool3d only support gpu version currently.
+    # RoIAwarePool3d only support gpu version currently.
+    if not torch.cuda.is_available():
         pytest.skip('test requires GPU and torch+cuda')
     roiaware_pool3d_max = RoIAwarePool3d(
         out_size=4, max_pts_per_voxel=128, mode='max')

tests/test_semantic_heads.py

+import pytest
+import torch
+
+
+def test_PointwiseSemanticHead():
+    # PointwiseSemanticHead only support gpu version currently.
+    if not torch.cuda.is_available():
+        pytest.skip('test requires GPU and torch+cuda')
+    from mmdet3d.models.builder import build_head
+
+    head_cfg = dict(
+        type='PointwiseSemanticHead',
+        in_channels=8,
+        extra_width=0.2,
+        seg_score_thr=0.3,
+        num_classes=3,
+        loss_seg=dict(
+            type='FocalLoss',
+            use_sigmoid=True,
+            reduction='sum',
+            gamma=2.0,
+            alpha=0.25,
+            loss_weight=1.0),
+        loss_part=dict(
+            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0))
+
+    self = build_head(head_cfg)
+    self.cuda()
+
+    # test forward
+    voxel_features = torch.rand([4, 8], dtype=torch.float32).cuda()
+    feats_dict = self.forward(voxel_features)
+    assert feats_dict['seg_preds'].shape == torch.Size(
+        [voxel_features.shape[0], 1])
+    assert feats_dict['part_preds'].shape == torch.Size(
+        [voxel_features.shape[0], 3])
+    assert feats_dict['part_feats'].shape == torch.Size(
+        [voxel_features.shape[0], 4])
+
+    voxel_centers = torch.tensor(
+        [[6.56126, 0.9648336, -1.7339306], [6.8162713, -2.480431, -1.3616394],
+         [11.643568, -4.744306, -1.3580885], [23.482342, 6.5036807, 0.5806964]
+         ],
+        dtype=torch.float32).cuda()  # n, point_features
+    coordinates = torch.tensor(
+        [[0, 12, 819, 131], [0, 16, 750, 136], [1, 16, 705, 232],
+         [1, 35, 930, 469]],
+        dtype=torch.int32).cuda()  # n, 4(batch, ind_x, ind_y, ind_z)
+    voxel_dict = dict(voxel_centers=voxel_centers, coors=coordinates)
+    gt_bboxes = list(
+        torch.tensor(
+            [[[6.4118, -3.4305, -1.7291, 1.7033, 3.4693, 1.6197, -0.9091]],
+             [[16.9107, 9.7925, -1.9201, 1.6097, 3.2786, 1.5307, -2.4056]]],
+            dtype=torch.float32).cuda())
+    gt_labels = list(torch.tensor([[0], [1]], dtype=torch.int64).cuda())
+
+    # test get_targets
+    target_dict = self.get_targets(voxel_dict, gt_bboxes, gt_labels)
+    assert target_dict['seg_targets'].shape == torch.Size(
+        [voxel_features.shape[0]])
+    assert target_dict['part_targets'].shape == torch.Size(
+        [voxel_features.shape[0], 3])
+
+    # test loss
+    loss_dict = self.loss(feats_dict['seg_preds'], feats_dict['part_preds'],
+                          target_dict['seg_targets'],
+                          target_dict['part_targets'])
+    assert loss_dict['loss_seg'] > 0
+    assert loss_dict['loss_part'] == 0  # no points in gt_boxes
+    total_loss = loss_dict['loss_seg'] + loss_dict['loss_part']
+    total_loss.backward()
+
+
+if __name__ == '__main__':
+    test_PointwiseSemanticHead()