[Feature] FCOS3D BBox Coder (#940)

* FCOS3D BBox Coder

* Add unit tests

* Change the value from long to float/double

* Rename bbox_out as bbox

* Add comments to forward returns

configs/_base_/models/fcos3d.py

@@ -55,6 +55,7 @@ model = dict(
             type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
         loss_centerness=dict(
             type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+        bbox_coder=dict(type='FCOS3DBBoxCoder', code_size=9),
         norm_on_bbox=True,
         centerness_on_reg=True,
         center_sampling=True,

mmdet3d/core/bbox/coders/__init__.py

@@ -3,6 +3,7 @@ from mmdet.core.bbox import build_bbox_coder
 from .anchor_free_bbox_coder import AnchorFreeBBoxCoder
 from .centerpoint_bbox_coders import CenterPointBBoxCoder
 from .delta_xyzwhlr_bbox_coder import DeltaXYZWLHRBBoxCoder
+from .fcos3d_bbox_coder import FCOS3DBBoxCoder
 from .groupfree3d_bbox_coder import GroupFree3DBBoxCoder
 from .partial_bin_based_bbox_coder import PartialBinBasedBBoxCoder
 from .point_xyzwhlr_bbox_coder import PointXYZWHLRBBoxCoder
@@ -10,5 +11,5 @@ from .point_xyzwhlr_bbox_coder import PointXYZWHLRBBoxCoder
 __all__ = [
     'build_bbox_coder', 'DeltaXYZWLHRBBoxCoder', 'PartialBinBasedBBoxCoder',
     'CenterPointBBoxCoder', 'AnchorFreeBBoxCoder', 'GroupFree3DBBoxCoder',
-    'PointXYZWHLRBBoxCoder'
+    'PointXYZWHLRBBoxCoder', 'FCOS3DBBoxCoder'
 ]

mmdet3d/core/bbox/coders/fcos3d_bbox_coder.py

+import numpy as np
+import torch
+
+from mmdet.core.bbox import BaseBBoxCoder
+from mmdet.core.bbox.builder import BBOX_CODERS
+from ..structures import limit_period
+
+
+@BBOX_CODERS.register_module()
+class FCOS3DBBoxCoder(BaseBBoxCoder):
+    """Bounding box coder for FCOS3D.
+
+    Args:
+        base_depths (tuple[tuple[float]]): Depth references for decode box
+            depth. Defaults to None.
+        base_dims (tuple[tuple[float]]): Dimension references for decode box
+            dimension. Defaults to None.
+        code_size (int): The dimension of boxes to be encoded. Defaults to 7.
+        norm_on_bbox (bool): Whether to apply normalization on the bounding
+            box 2D attributes. Defaults to True.
+    """
+
+    def __init__(self,
+                 base_depths=None,
+                 base_dims=None,
+                 code_size=7,
+                 norm_on_bbox=True):
+        super(FCOS3DBBoxCoder, self).__init__()
+        self.base_depths = base_depths
+        self.base_dims = base_dims
+        self.bbox_code_size = code_size
+        self.norm_on_bbox = norm_on_bbox
+
+    def encode(self, gt_bboxes_3d, gt_labels_3d, gt_bboxes, gt_labels):
+        # TODO: refactor the encoder in the FCOS3D and PGD head
+        pass
+
+    def decode(self, bbox, scale, stride, training, cls_score=None):
+        """Decode regressed results into 3D predictions.
+
+        Note that offsets are not transformed to the projected 3D centers.
+
+        Args:
+            bbox (torch.Tensor): Raw bounding box predictions in shape
+                [N, C, H, W].
+            scale (tuple[`Scale`]): Learnable scale parameters.
+            stride (tuple[int]): Stride for a specific feature level.
+            training (bool): Whether the decoding is in the training
+                procedure.
+            cls_score (torch.Tensor): Classification score map for deciding
+                which base depth or dim is used. Defaults to None.
+
+        Returns:
+            torch.Tensor: Decoded boxes.
+        """
+        # scale the bbox of different level
+        # only apply to offset, depth and size prediction
+        scale_offset, scale_depth, scale_size = scale[0:3]
+
+        clone_bbox = bbox.clone()
+        bbox[:, :2] = scale_offset(clone_bbox[:, :2]).float()
+        bbox[:, 2] = scale_depth(clone_bbox[:, 2]).float()
+        bbox[:, 3:6] = scale_size(clone_bbox[:, 3:6]).float()
+
+        if self.base_depths is None:
+            bbox[:, 2] = bbox[:, 2].exp()
+        elif len(self.base_depths) == 1:  # only single prior
+            mean = self.base_depths[0][0]
+            std = self.base_depths[0][1]
+            bbox[:, 2] = mean + bbox.clone()[:, 2] * std
+        else:  # multi-class priors
+            assert len(self.base_depths) == cls_score.shape[1], \
+                'The number of multi-class depth priors should be equal to ' \
+                'the number of categories.'
+            indices = cls_score.max(dim=1)[1]
+            depth_priors = cls_score.new_tensor(
+                self.base_depths)[indices, :].permute(0, 3, 1, 2)
+            mean = depth_priors[:, 0]
+            std = depth_priors[:, 1]
+            bbox[:, 2] = mean + bbox.clone()[:, 2] * std
+
+        bbox[:, 3:6] = bbox[:, 3:6].exp()
+        if self.base_dims is not None:
+            assert len(self.base_dims) == cls_score.shape[1], \
+                'The number of anchor sizes should be equal to the number ' \
+                'of categories.'
+            indices = cls_score.max(dim=1)[1]
+            size_priors = cls_score.new_tensor(
+                self.base_dims)[indices, :].permute(0, 3, 1, 2)
+            bbox[:, 3:6] = size_priors * bbox.clone()[:, 3:6]
+
+        assert self.norm_on_bbox is True, 'Setting norm_on_bbox to False '\
+            'has not been thoroughly tested for FCOS3D.'
+        if self.norm_on_bbox:
+            if not training:
+                # Note that this line is conducted only when testing
+                bbox[:, :2] *= stride
+
+        return bbox
+
+    @staticmethod
+    def decode_yaw(bbox, centers2d, dir_cls, dir_offset, cam2img):
+        """Decode yaw angle and change it from local to global.i.
+
+        Args:
+            bbox (torch.Tensor): Bounding box predictions in shape
+                [N, C] with yaws to be decoded.
+            centers2d (torch.Tensor): Projected 3D-center on the image planes
+                corresponding to the box predictions.
+            dir_cls (torch.Tensor): Predicted direction classes.
+            dir_offset (float): Direction offset before dividing all the
+                directions into several classes.
+            cam2img (torch.Tensor): Camera intrinsic matrix in shape [4, 4].
+
+        Returns:
+            torch.Tensor: Bounding boxes with decoded yaws.
+        """
+        if bbox.shape[0] > 0:
+            dir_rot = limit_period(bbox[..., 6] - dir_offset, 0, np.pi)
+            bbox[..., 6] = \
+                dir_rot + dir_offset + np.pi * dir_cls.to(bbox.dtype)
+
+        bbox[:, 6] = torch.atan2(centers2d[:, 0] - cam2img[0, 2],
+                                 cam2img[0, 0]) + bbox[:, 6]
+
+        return bbox

mmdet3d/models/dense_heads/fcos_mono3d_head.py

@@ -9,6 +9,7 @@ from torch import nn as nn
 from mmdet3d.core import (box3d_multiclass_nms, limit_period, points_img2cam,
                           xywhr2xyxyr)
 from mmdet.core import multi_apply
+from mmdet.core.bbox.builder import build_bbox_coder
 from mmdet.models.builder import HEADS, build_loss
 from .anchor_free_mono3d_head import AnchorFreeMono3DHead
 
@@ -75,6 +76,7 @@ class FCOSMono3DHead(AnchorFreeMono3DHead):
                      type='CrossEntropyLoss',
                      use_sigmoid=True,
                      loss_weight=1.0),
+                 bbox_coder=dict(type='FCOS3DBBoxCoder', code_size=9),
                  norm_cfg=dict(type='GN', num_groups=32, requires_grad=True),
                  centerness_branch=(64, ),
                  init_cfg=None,
@@ -97,6 +99,8 @@ class FCOSMono3DHead(AnchorFreeMono3DHead):
             init_cfg=init_cfg,
             **kwargs)
         self.loss_centerness = build_loss(loss_centerness)
+        bbox_coder['code_size'] = self.bbox_code_size
+        self.bbox_coder = build_bbox_coder(bbox_coder)
         if init_cfg is None:
             self.init_cfg = dict(
                 type='Normal',
@@ -112,9 +116,11 @@ class FCOSMono3DHead(AnchorFreeMono3DHead):
             conv_channels=self.centerness_branch,
             conv_strides=(1, ) * len(self.centerness_branch))
         self.conv_centerness = nn.Conv2d(self.centerness_branch[-1], 1, 1)
+        self.scale_dim = 3  # only for offset, depth and size regression
         self.scales = nn.ModuleList([
-            nn.ModuleList([Scale(1.0) for _ in range(3)]) for _ in self.strides
-        ])  # only for offset, depth and size regression
+            nn.ModuleList([Scale(1.0) for _ in range(self.scale_dim)])
+            for _ in self.strides
+        ])
 
     def forward(self, feats):
         """Forward features from the upstream network.
@@ -140,8 +146,9 @@ class FCOSMono3DHead(AnchorFreeMono3DHead):
                 centernesses (list[Tensor]): Centerness for each scale level,
                     each is a 4D-tensor, the channel number is num_points * 1.
         """
+        # Note: we use [:5] to filter feats and only return predictions
         return multi_apply(self.forward_single, feats, self.scales,
-                           self.strides)
+                           self.strides)[:5]
 
     def forward_single(self, x, scale, stride):
         """Forward features of a single scale levle.
@@ -171,26 +178,12 @@ class FCOSMono3DHead(AnchorFreeMono3DHead):
             for conv_centerness_prev_layer in self.conv_centerness_prev:
                 clone_cls_feat = conv_centerness_prev_layer(clone_cls_feat)
             centerness = self.conv_centerness(clone_cls_feat)
-        # scale the bbox_pred of different level
-        # only apply to offset, depth and size prediction
-        scale_offset, scale_depth, scale_size = scale[0:3]
-
-        clone_bbox_pred = bbox_pred.clone()
-        bbox_pred[:, :2] = scale_offset(clone_bbox_pred[:, :2]).float()
-        bbox_pred[:, 2] = scale_depth(clone_bbox_pred[:, 2]).float()
-        bbox_pred[:, 3:6] = scale_size(clone_bbox_pred[:, 3:6]).float()
 
-        bbox_pred[:, 2] = bbox_pred[:, 2].exp()
-        bbox_pred[:, 3:6] = bbox_pred[:, 3:6].exp() + 1e-6  # avoid size=0
+        bbox_pred = self.bbox_coder.decode(bbox_pred, scale, stride,
+                                           self.training, cls_score)
 
-        assert self.norm_on_bbox is True, 'Setting norm_on_bbox to False '\
-            'has not been thoroughly tested for FCOS3D.'
-        if self.norm_on_bbox:
-            if not self.training:
-                # Note that this line is conducted only when testing
-                bbox_pred[:, :2] *= stride
-
-        return cls_score, bbox_pred, dir_cls_pred, attr_pred, centerness
+        return cls_score, bbox_pred, dir_cls_pred, attr_pred, centerness, \
+            cls_feat, reg_feat
 
     @staticmethod
     def add_sin_difference(boxes1, boxes2):
@@ -654,19 +647,13 @@ class FCOSMono3DHead(AnchorFreeMono3DHead):
         mlvl_dir_scores = torch.cat(mlvl_dir_scores)
 
         # change local yaw to global yaw for 3D nms
-        if mlvl_bboxes.shape[0] > 0:
-            dir_rot = limit_period(mlvl_bboxes[..., 6] - self.dir_offset, 0,
-                                   np.pi)
-            mlvl_bboxes[..., 6] = (
-                dir_rot + self.dir_offset +
-                np.pi * mlvl_dir_scores.to(mlvl_bboxes.dtype))
-
-        cam_intrinsic = mlvl_centers2d.new_zeros((4, 4))
-        cam_intrinsic[:view.shape[0], :view.shape[1]] = \
+        cam2img = mlvl_centers2d.new_zeros((4, 4))
+        cam2img[:view.shape[0], :view.shape[1]] = \
             mlvl_centers2d.new_tensor(view)
-        mlvl_bboxes[:, 6] = torch.atan2(
-            mlvl_centers2d[:, 0] - cam_intrinsic[0, 2],
-            cam_intrinsic[0, 0]) + mlvl_bboxes[:, 6]
+        mlvl_bboxes = self.bbox_coder.decode_yaw(mlvl_bboxes, mlvl_centers2d,
+                                                 mlvl_dir_scores,
+                                                 self.dir_offset, cam2img)
+
         mlvl_bboxes_for_nms = xywhr2xyxyr(input_meta['box_type_3d'](
             mlvl_bboxes, box_dim=self.bbox_code_size,
             origin=(0.5, 0.5, 0.5)).bev)

tests/test_utils/test_bbox_coders.py

 # Copyright (c) OpenMMLab. All rights reserved.
 import torch
+from mmcv.cnn import Scale
+from torch import nn as nn
 
 from mmdet3d.core.bbox import DepthInstance3DBoxes, LiDARInstance3DBoxes
 from mmdet.core import build_bbox_coder
@@ -382,3 +384,84 @@ def test_point_xyzwhlr_bbox_coder():
     # test decode
     bbox3d_out = boxcoder.decode(bbox_target, points, gt_labels_3d)
     assert torch.allclose(bbox3d_out, gt_bboxes_3d, atol=1e-4)
+
+
+def test_fcos3d_bbox_coder():
+    # test a config without priors
+    bbox_coder_cfg = dict(
+        type='FCOS3DBBoxCoder',
+        base_depths=None,
+        base_dims=None,
+        code_size=7,
+        norm_on_bbox=True)
+    bbox_coder = build_bbox_coder(bbox_coder_cfg)
+
+    # test decode
+    # [2, 7, 1, 1]
+    batch_bbox_out = torch.tensor([[[[0.3130]], [[0.7094]], [[0.8743]],
+                                    [[0.0570]], [[0.5579]], [[0.1593]],
+                                    [[0.4553]]],
+                                   [[[0.7758]], [[0.2298]], [[0.3925]],
+                                    [[0.6307]], [[0.4377]], [[0.3339]],
+                                    [[0.1966]]]])
+    batch_scale = nn.ModuleList([Scale(1.0) for _ in range(3)])
+    stride = 2
+    training = False
+    cls_score = torch.randn([2, 2, 1, 1]).sigmoid()
+    decode_bbox_out = bbox_coder.decode(batch_bbox_out, batch_scale, stride,
+                                        training, cls_score)
+
+    expected_bbox_out = torch.tensor([[[[0.6261]], [[1.4188]], [[2.3971]],
+                                       [[1.0586]], [[1.7470]], [[1.1727]],
+                                       [[0.4553]]],
+                                      [[[1.5516]], [[0.4596]], [[1.4806]],
+                                       [[1.8790]], [[1.5492]], [[1.3965]],
+                                       [[0.1966]]]])
+    assert torch.allclose(decode_bbox_out, expected_bbox_out, atol=1e-3)
+
+    # test a config with priors
+    prior_bbox_coder_cfg = dict(
+        type='FCOS3DBBoxCoder',
+        base_depths=((28., 13.), (25., 12.)),
+        base_dims=((2., 3., 1.), (1., 2., 3.)),
+        code_size=7,
+        norm_on_bbox=True)
+    prior_bbox_coder = build_bbox_coder(prior_bbox_coder_cfg)
+
+    # test decode
+    batch_bbox_out = torch.tensor([[[[0.3130]], [[0.7094]], [[0.8743]],
+                                    [[0.0570]], [[0.5579]], [[0.1593]],
+                                    [[0.4553]]],
+                                   [[[0.7758]], [[0.2298]], [[0.3925]],
+                                    [[0.6307]], [[0.4377]], [[0.3339]],
+                                    [[0.1966]]]])
+    batch_scale = nn.ModuleList([Scale(1.0) for _ in range(3)])
+    stride = 2
+    training = False
+    cls_score = torch.tensor([[[[0.5811]], [[0.6198]]], [[[0.4889]],
+                                                         [[0.8142]]]])
+    decode_bbox_out = prior_bbox_coder.decode(batch_bbox_out, batch_scale,
+                                              stride, training, cls_score)
+    expected_bbox_out = torch.tensor([[[[0.6260]], [[1.4188]], [[35.4916]],
+                                       [[1.0587]], [[3.4940]], [[3.5181]],
+                                       [[0.4553]]],
+                                      [[[1.5516]], [[0.4596]], [[29.7100]],
+                                       [[1.8789]], [[3.0983]], [[4.1892]],
+                                       [[0.1966]]]])
+    assert torch.allclose(decode_bbox_out, expected_bbox_out, atol=1e-3)
+
+    # test decode_yaw
+    decode_bbox_out = decode_bbox_out.permute(0, 2, 3, 1).view(-1, 7)
+    batch_centers2d = torch.tensor([[100., 150.], [200., 100.]])
+    batch_dir_cls = torch.tensor([0., 1.])
+    dir_offset = 0.7854
+    cam2img = torch.tensor([[700., 0., 450., 0.], [0., 700., 200., 0.],
+                            [0., 0., 1., 0.], [0., 0., 0., 1.]])
+    decode_bbox_out = prior_bbox_coder.decode_yaw(decode_bbox_out,
+                                                  batch_centers2d,
+                                                  batch_dir_cls, dir_offset,
+                                                  cam2img)
+    expected_bbox_out = torch.tensor(
+        [[0.6260, 1.4188, 35.4916, 1.0587, 3.4940, 3.5181, 3.1332],
+         [1.5516, 0.4596, 29.7100, 1.8789, 3.0983, 4.1892, 6.1368]])
+    assert torch.allclose(decode_bbox_out, expected_bbox_out, atol=1e-3)