Change cam rot_3d_in_axis (#906)

mmdet3d/core/bbox/structures/cam_box3d.py

@@ -137,11 +137,8 @@ class CameraInstance3DBoxes(BaseInstance3DBoxes):
         corners_norm = corners_norm - dims.new_tensor([0.5, 1, 0.5])
         corners = dims.view([-1, 1, 3]) * corners_norm.reshape([1, 8, 3])
 
-        # positive direction of the gravity axis
-        # in cam coord system points to the earth
-        # so the rotation is clockwise if viewed from above
         corners = rotation_3d_in_axis(
-            corners, self.tensor[:, 6], axis=self.YAW_AXIS, clockwise=True)
+            corners, self.tensor[:, 6], axis=self.YAW_AXIS)
         corners += self.tensor[:, :3].view(-1, 1, 3)
         return corners
 
@@ -182,11 +179,7 @@ class CameraInstance3DBoxes(BaseInstance3DBoxes):
                 self.tensor[:, 0:3],
                 angle,
                 axis=self.YAW_AXIS,
-                return_mat=True,
-                # positive direction of the gravity axis
-                # in cam coord system points to the earth
-                # so the rotation is clockwise if viewed from above
-                clockwise=True)
+                return_mat=True)
         else:
             rot_mat_T = angle
             rot_sin = rot_mat_T[2, 0]

mmdet3d/core/bbox/structures/utils.py

@@ -71,9 +71,9 @@ def rotation_3d_in_axis(points,
     if points.shape[-1] == 3:
         if axis == 1 or axis == -2:
             rot_mat_T = torch.stack([
-                torch.stack([rot_cos, zeros, rot_sin]),
+                torch.stack([rot_cos, zeros, -rot_sin]),
                 torch.stack([zeros, ones, zeros]),
-                torch.stack([-rot_sin, zeros, rot_cos])
+                torch.stack([rot_sin, zeros, rot_cos])
             ])
         elif axis == 2 or axis == -1:
             rot_mat_T = torch.stack([

mmdet3d/models/dense_heads/ssd_3d_head.py

@@ -391,7 +391,8 @@ class SSD3DHead(VoteHead):
             # LiDARInstance3DBoxes and DepthInstance3DBoxes
             canonical_xyz = rotation_3d_in_axis(
                 canonical_xyz.unsqueeze(0).transpose(0, 1),
-                -gt_bboxes_3d.yaw[assignment], 2).squeeze(1)
+                -gt_bboxes_3d.yaw[assignment],
+                axis=2).squeeze(1)
         distance_front = torch.clamp(
             size_res_targets[:, 0] - canonical_xyz[:, 0], min=0)
         distance_back = torch.clamp(

tests/test_models/test_detectors.py

@@ -437,7 +437,8 @@ def test_imvoxelnet():
     if not torch.cuda.is_available():
         pytest.skip('test requires GPU and torch+cuda')
 
-    imvoxelnet_cfg = _get_detector_cfg('imvoxelnet/imvoxelnet_kitti-3d-car.py')
+    imvoxelnet_cfg = _get_detector_cfg(
+        'imvoxelnet/imvoxelnet_4x8_kitti-3d-car.py')
     self = build_detector(imvoxelnet_cfg).cuda()
     imgs = torch.rand([1, 3, 384, 1280], dtype=torch.float32).cuda()
     gt_bboxes_3d = [LiDARInstance3DBoxes(torch.rand([3, 7], device='cuda'))]

tests/test_utils/test_box3d.py

@@ -1553,13 +1553,37 @@ def test_rotation_3d_in_axis():
                            [[-0.2555, -0.2683, 0.0000],
                             [-0.2555, -0.2683, 0.9072]]])
     angles = [np.pi / 2, -np.pi / 2]
-    rotated = rotation_3d_in_axis(points, angles, axis=0)
+    rotated = rotation_3d_in_axis(points, angles, axis=0).numpy()
     expected_rotated = np.array([[[-0.4599, 0.0000, -0.0471],
                                   [-0.4599, -1.8433, -0.0471]],
                                  [[-0.2555, 0.0000, 0.2683],
                                   [-0.2555, 0.9072, 0.2683]]])
     assert np.allclose(rotated, expected_rotated, atol=1e-3)
 
+    points = torch.tensor([[[-0.4599, -0.0471, 0.0000],
+                            [-0.4599, -0.0471, 1.8433]],
+                           [[-0.2555, -0.2683, 0.0000],
+                            [-0.2555, -0.2683, 0.9072]]])
+    angles = [np.pi / 2, -np.pi / 2]
+    rotated = rotation_3d_in_axis(points, angles, axis=1).numpy()
+    expected_rotated = np.array([[[0.0000, -0.0471, 0.4599],
+                                  [1.8433, -0.0471, 0.4599]],
+                                 [[0.0000, -0.2683, -0.2555],
+                                  [-0.9072, -0.2683, -0.2555]]])
+    assert np.allclose(rotated, expected_rotated, atol=1e-3)
+
+    points = torch.tensor([[[-0.4599, -0.0471, 0.0000],
+                            [-0.4599, 0.0471, 1.8433]],
+                           [[-0.2555, -0.2683, 0.0000],
+                            [0.2555, -0.2683, 0.9072]]])
+    angles = [np.pi / 2, -np.pi / 2]
+    rotated = rotation_3d_in_axis(points, angles, axis=2).numpy()
+    expected_rotated = np.array([[[0.0471, -0.4599, 0.0000],
+                                  [-0.0471, -0.4599, 1.8433]],
+                                 [[-0.2683, 0.2555, 0.0000],
+                                  [-0.2683, -0.2555, 0.9072]]])
+    assert np.allclose(rotated, expected_rotated, atol=1e-3)
+
     points = torch.tensor([[[-0.0471, 0.0000], [-0.0471, 1.8433]],
                            [[-0.2683, 0.0000], [-0.2683, 0.9072]]])
     angles = [np.pi / 2, -np.pi / 2]