[Fix] Fix the dimension and yaw hack of FCOS3D on nuScenes

configs/fcos3d/README.md

@@ -58,6 +58,6 @@ We also provide visualization functions to show the monocular 3D detection resul
 
 |  Backbone   | Lr schd | Mem (GB) | Inf time (fps) | mAP | NDS | Download |
 | :---------: | :-----: | :------: | :------------: | :----: |:----: | :------: |
-|[ResNet101 w/ DCN](./fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d.py)|1x|8.69||29.9|37.3|[model](https://download.openmmlab.com/mmdetection3d/v0.1.0_models/fcos3d/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_20210425_181341-8d5a21fe.pth) | [log](https://download.openmmlab.com/mmdetection3d/v0.1.0_models/fcos3d/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_20210425_181341.log.json)|
-|[above w/ finetune](./fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_finetune.py)|1x|8.69||32.1|39.3|[model](https://download.openmmlab.com/mmdetection3d/v0.1.0_models/fcos3d/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_finetune/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_finetune_20210427_091419-35aaaad0.pth) | [log](https://download.openmmlab.com/mmdetection3d/v0.1.0_models/fcos3d/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_finetune/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_finetune_20210427_091419.log.json)|
-|above w/ tta|1x|8.69||33.1|40.0||
+|[ResNet101 w/ DCN](./fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d.py)|1x|8.69||29.8|37.7|[model](https://download.openmmlab.com/mmdetection3d/v0.1.0_models/fcos3d/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_20210715_235813-4bed5239.pth) | [log](https://download.openmmlab.com/mmdetection3d/v0.1.0_models/fcos3d/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_20210715_235813.log.json)|
+|[above w/ finetune](./fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_finetune.py)|1x|8.69||32.1|39.5|[model](https://download.openmmlab.com/mmdetection3d/v0.1.0_models/fcos3d/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_finetune/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_finetune_20210717_095645-8d806dc2.pth) | [log](https://download.openmmlab.com/mmdetection3d/v0.1.0_models/fcos3d/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_finetune/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_finetune_20210717_095645.log.json)|
+|above w/ tta|1x|8.69||33.1|40.3||

docs/compatibility.md

@@ -4,6 +4,12 @@ This document provides detailed descriptions of the BC-breaking changes in MMDet
 
 ## MMDetection3D 0.16.0
 
+### NuScenes coco-style data pre-processing
+
+We remove the rotation and dimension hack in the monocular 3D detection on nuScenes. Specifically, we transform the rotation and dimension of boxes defined by nuScenes devkit to the coordinate system of our `CameraInstance3DBoxes` in the pre-processing and transform them back in the post-processing. In this way, we can remove the corresponding [hack](https://github.com/open-mmlab/mmdetection3d/pull/744/files#diff-5bee5062bd84e6fa25a2fdd71353f6f283dfdc4a66a0316c3b1ca26078c978b6L165) used in the visualization tools. The modification also guarantees the correctness of all the operations based on our `CameraInstance3DBoxes` (such as NMS and flip augmentation) when training monocular 3D detectors.
+
+The modification only influences nuScenes coco-style json files. Please re-run the nuScenes data preparation script if necessary. See more details in the PR [#744](https://github.com/open-mmlab/mmdetection3d/pull/744).
+
 ### ScanNet dataset for ImVoxelNet
 
 We adopt a new pre-processing procedure for the ScanNet dataset in order to support ImVoxelNet, which is a multi-view method requiring image data. In previous versions of MMDetection3D, ScanNet dataset was only used for point cloud based 3D detection and segmentation methods. We plan adding ImVoxelNet to our model zoo, thus updating ScanNet correspondingly by adding image-related pre-processing steps. Specifically, we made these changes:
@@ -22,12 +28,12 @@ In order to fix the problem that the priority of EvalHook is too low, all hook p
 
 ### Unified parameter initialization
 
-To unify the parameter initialization in OpenMMLab projects, MMCV supports `BaseModule` that accepts `init_cfg` to allow the modules' parameters initialized in a flexible and unified manner. Now the users need to explicitly call `model.init_weights()` in the training script to initialize the model (as in [here](https://github.com/open-mmlab/mmdetection3d/blob/master/tools/train.py#L183), previously this was handled by the detector. Please refer to PR #622 for details.
+To unify the parameter initialization in OpenMMLab projects, MMCV supports `BaseModule` that accepts `init_cfg` to allow the modules' parameters initialized in a flexible and unified manner. Now the users need to explicitly call `model.init_weights()` in the training script to initialize the model (as in [here](https://github.com/open-mmlab/mmdetection3d/blob/master/tools/train.py#L183), previously this was handled by the detector. Please refer to PR [#622](https://github.com/open-mmlab/mmdetection3d/pull/622) for details.
 
 ### BackgroundPointsFilter
 
-We modified the dataset aumentation function `BackgroundPointsFilter`(in [here](https://github.com/open-mmlab/mmdetection3d/blob/mmdet3d/datasets/pipelines/transforms_3d.py#L1101)). In previous version of MMdetection3D, `BackgroundPointsFilter` changes the gt_bboxes_3d's bottom center to the gravity center. In MMDetection3D 0.15.0,
-`BackgroundPointsFilter` will not change it. Please refer to PR #609 for details.
+We modified the dataset augmentation function `BackgroundPointsFilter`([here](https://github.com/open-mmlab/mmdetection3d/blob/v0.15.0/mmdet3d/datasets/pipelines/transforms_3d.py#L1132)). In previous version of MMdetection3D, `BackgroundPointsFilter` changes the gt_bboxes_3d's bottom center to the gravity center. In MMDetection3D 0.15.0,
+`BackgroundPointsFilter` will not change it. Please refer to PR [#609](https://github.com/open-mmlab/mmdetection3d/pull/609) for details.
 
 ### Enhance `IndoorPatchPointSample` transform
 
@@ -57,7 +63,7 @@ We have trained a [VoteNet](https://github.com/open-mmlab/mmdetection3d/blob/mas
 
 ### SUNRGBD dataset for ImVoteNet
 
-We adopt a new pre-processing procedure for the SUNRGBD dataset in order to support ImVoteNet, which is a multi-modality method requiring both image and point cloud data. In previous versions of MMDetection3D, SUNRGBD dataset was only used for point cloud based 3D detection methods. In MMDetection3D 0.12.0, we add ImVoteNet to our model zoo, thus updating SUNRGBD correspondingly by adding image-related pre-processing steps. Specificly, we made these changes:
+We adopt a new pre-processing procedure for the SUNRGBD dataset in order to support ImVoteNet, which is a multi-modality method requiring both image and point cloud data. In previous versions of MMDetection3D, SUNRGBD dataset was only used for point cloud based 3D detection methods. In MMDetection3D 0.12.0, we add ImVoteNet to our model zoo, thus updating SUNRGBD correspondingly by adding image-related pre-processing steps. Specifically, we made these changes:
 
 - Fix a bug in the image file path in meta data.
 - Convert calibration matrices from double to float to avoid type mismatch in further operations.

docs/demo.md

@@ -70,6 +70,8 @@ Example on nuScenes data using [FCOS3D](https://github.com/open-mmlab/mmdetectio
 python demo/mono_det_demo.py demo/data/nuscenes/n015-2018-07-24-11-22-45+0800__CAM_BACK__1532402927637525.jpg demo/data/nuscenes/n015-2018-07-24-11-22-45+0800__CAM_BACK__1532402927637525_mono3d.coco.json configs/fcos3d/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_finetune.py checkpoints/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_finetune_20210427_091419-35aaaad0.pth
 ```
 
+Note that when visualizing results of monocular 3D detection for flipped images, the camera intrinsic matrix should also be modified accordingly. See more details and examples in PR [#744](https://github.com/open-mmlab/mmdetection3d/pull/744).
+
 ### 3D Segmentation
 
 To test a 3D segmentor on point cloud data, simply run:

mmdet3d/datasets/kitti_mono_dataset.py

@@ -88,9 +88,6 @@ def _parse_ann_info(self, img_info, ann_info):
                 gt_masks_ann.append(ann.get('segmentation', None))
                 # 3D annotations in camera coordinates
                 bbox_cam3d = np.array(ann['bbox_cam3d']).reshape(-1, )
-                # change orientation to local yaw
-                bbox_cam3d[6] = -np.arctan2(bbox_cam3d[0],
-                                            bbox_cam3d[2]) + bbox_cam3d[6]
                 gt_bboxes_cam3d.append(bbox_cam3d)
                 # 2.5D annotations in camera coordinates
                 center2d = ann['center2d'][:2]

mmdet3d/datasets/nuscenes_mono_dataset.py

@@ -173,9 +173,6 @@ def _parse_ann_info(self, img_info, ann_info):
                 gt_masks_ann.append(ann.get('segmentation', None))
                 # 3D annotations in camera coordinates
                 bbox_cam3d = np.array(ann['bbox_cam3d']).reshape(1, -1)
-                # change orientation to local yaw
-                bbox_cam3d[0, 6] = -np.arctan2(
-                    bbox_cam3d[0, 0], bbox_cam3d[0, 2]) + bbox_cam3d[0, 6]
                 velo_cam3d = np.array(ann['velo_cam3d']).reshape(1, 2)
                 nan_mask = np.isnan(velo_cam3d[:, 0])
                 velo_cam3d[nan_mask] = [0.0, 0.0]
@@ -666,6 +663,10 @@ def output_to_nusc_box(detection):
     box_dims = box3d.dims.numpy()
     box_yaw = box3d.yaw.numpy()
 
+    # convert the dim/rot to nuscbox convention
+    box_dims[:, [0, 1, 2]] = box_dims[:, [2, 0, 1]]
+    box_yaw = -box_yaw
+
     box_list = []
     for i in range(len(box3d)):
         q1 = pyquaternion.Quaternion(axis=[0, 0, 1], radians=box_yaw[i])
@@ -778,6 +779,11 @@ def nusc_box_to_cam_box3d(boxes):
     rots = torch.Tensor([b.orientation.yaw_pitch_roll[0]
                          for b in boxes]).view(-1, 1)
     velocity = torch.Tensor([b.velocity[:2] for b in boxes]).view(-1, 2)
+
+    # convert nusbox to cambox convention
+    dims[:, [0, 1, 2]] = dims[:, [1, 2, 0]]
+    rots = -rots
+
     boxes_3d = torch.cat([locs, dims, rots, velocity], dim=1).cuda()
     cam_boxes3d = CameraInstance3DBoxes(
         boxes_3d, box_dim=9, origin=(0.5, 0.5, 0.5))

mmdet3d/datasets/pipelines/transforms_3d.py

@@ -119,9 +119,16 @@ def random_flip_data_3d(self, input_dict, direction='horizontal'):
         if 'centers2d' in input_dict:
             assert self.sync_2d is True and direction == 'horizontal', \
                 'Only support sync_2d=True and horizontal flip with images'
-            w = input_dict['img_shape'][1]
+            w = input_dict['ori_shape'][1]
             input_dict['centers2d'][..., 0] = \
                 w - input_dict['centers2d'][..., 0]
+            # need to modify the horizontal position of camera center
+            # along u-axis in the image (flip like centers2d)
+            # ['cam_intrinsic'][0][2] = c_u
+            # see more details and examples at
+            # https://github.com/open-mmlab/mmdetection3d/pull/744
+            input_dict['cam_intrinsic'][0][2] = \
+                w - input_dict['cam_intrinsic'][0][2]
 
     def __call__(self, input_dict):
         """Call function to flip points, values in the ``bbox3d_fields`` and \

mmdet3d/models/dense_heads/fcos_mono3d_head.py

@@ -861,6 +861,10 @@ def _get_target_single(self, gt_bboxes, gt_labels, gt_bboxes_3d,
                    attr_labels.new_full(
                        (num_points,), self.attr_background_label)
 
+        # change orientation to local yaw
+        gt_bboxes_3d[..., 6] = -torch.atan2(
+            gt_bboxes_3d[..., 0], gt_bboxes_3d[..., 2]) + gt_bboxes_3d[..., 6]
+
         areas = (gt_bboxes[:, 2] - gt_bboxes[:, 0]) * (
             gt_bboxes[:, 3] - gt_bboxes[:, 1])
         areas = areas[None].repeat(num_points, 1)

tools/data_converter/nuscenes_converter.py

@@ -483,8 +483,13 @@ def get_2d_boxes(nusc,
         # If mono3d=True, add 3D annotations in camera coordinates
         if mono3d and (repro_rec is not None):
             loc = box.center.tolist()
-            dim = box.wlh.tolist()
-            rot = [box.orientation.yaw_pitch_roll[0]]
+
+            dim = box.wlh
+            dim[[0, 1, 2]] = dim[[1, 2, 0]]  # convert wlh to our lhw
+            dim = dim.tolist()
+
+            rot = box.orientation.yaw_pitch_roll[0]
+            rot = [-rot]  # convert the rot to our cam coordinate
 
             global_velo2d = nusc.box_velocity(box.token)[:2]
             global_velo3d = np.array([*global_velo2d, 0.0])

tools/misc/browse_dataset.py

@@ -159,10 +159,6 @@ def show_proj_bbox_img(idx,
             img_metas=img_metas,
             show=show)
     elif isinstance(gt_bboxes, CameraInstance3DBoxes):
-        # TODO: remove the hack of box from NuScenesMonoDataset
-        if is_nus_mono:
-            from mmdet3d.core.bbox import mono_cam_box2vis
-            gt_bboxes = mono_cam_box2vis(gt_bboxes)
         show_multi_modality_result(
             img,
             gt_bboxes,