[Fix] Fix bg point filter (#609)

* fix bg_points_filter

* fix 3dssd bg_point_filter bug

* fix bg_filter bug in 3dssd

* update performance

* add model download link

configs/3dssd/3dssd_kitti-3d-car.py → configs/3dssd/3dssd_4x4_kitti-3d-car.py

@@ -106,9 +106,9 @@ model = dict(
 lr = 0.002  # max learning rate
 optimizer = dict(type='AdamW', lr=lr, weight_decay=0)
 optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))
-lr_config = dict(policy='step', warmup=None, step=[80, 120])
+lr_config = dict(policy='step', warmup=None, step=[45, 60])
 # runtime settings
-runner = dict(type='EpochBasedRunner', max_epochs=150)
+runner = dict(type='EpochBasedRunner', max_epochs=80)
 
 # yapf:disable
 log_config = dict(

configs/3dssd/README.md

@@ -29,6 +29,6 @@ Some settings in our implementation are different from the [official implementat
 
 |  Backbone   |Class| Lr schd | Mem (GB) | Inf time (fps) | mAP |Download |
 | :---------: | :-----: | :------: | :------------: | :----: |:----: | :------: |
-|    [PointNet2SAMSG](./3dssd_kitti-3d-car.py)| Car |72e|4.7||78.39(81.00)<sup>1</sup>|[model](https://download.openmmlab.com/mmdetection3d/v0.1.0_models/3dssd/3dssd_kitti-3d-car_20210324_122002-07e9a19b.pth) &#124; [log](https://download.openmmlab.com/mmdetection3d/v0.1.0_models/3dssd/3dssd_kitti-3d-car_20210324_122002.log.json)|
+|    [PointNet2SAMSG](./3dssd_kitti-3d-car.py)| Car |72e|4.7||78.69(81.27)<sup>1</sup>|[model](https://download.openmmlab.com/mmdetection3d/v0.1.0_models/3dssd/3dssd_kitti-3d-car_20210602_124438-b4276f56.pth) &#124; [log](https://download.openmmlab.com/mmdetection3d/v0.1.0_models/3dssd/3dssd_kitti-3d-car_20210602_124438.log.json)|
 
-[1]: We report two different 3D object detection performance here. 78.39mAP is evaluated by our evaluation code and 81.00mAP is evaluated by the official development kit （so as that used in the paper and official code of 3DSSD ）. We found that the commonly used Python implementation of [`rotate_iou`](https://github.com/traveller59/second.pytorch/blob/e42e4a0e17262ab7d180ee96a0a36427f2c20a44/second/core/non_max_suppression/nms_gpu.py#L605) which is used in our KITTI dataset evaluation, is different from the official implemention in [KITTI benchmark](http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d).
+[1]: We report two different 3D object detection performance here. 78.69mAP is evaluated by our evaluation code and 81.27mAP is evaluated by the official development kit （so as that used in the paper and official code of 3DSSD ）. We found that the commonly used Python implementation of [`rotate_iou`](https://github.com/traveller59/second.pytorch/blob/e42e4a0e17262ab7d180ee96a0a36427f2c20a44/second/core/non_max_suppression/nms_gpu.py#L605) which is used in our KITTI dataset evaluation, is different from the official implemention in [KITTI benchmark](http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d).

configs/3dssd/metafile.yml

@@ -20,5 +20,5 @@ Models:
       - Task: 3D Object Detection
         Dataset: KITTI
         Metrics:
-          mAP: 78.39
-    Weights: https://download.openmmlab.com/mmdetection3d/v0.1.0_models/3dssd/3dssd_kitti-3d-car_20210324_122002-07e9a19b.pth
+          mAP: 78.69
+    Weights: https://download.openmmlab.com/mmdetection3d/v0.1.0_models/3dssd/3dssd_kitti-3d-car_20210602_124438-b4276f56.pth

mmdet3d/datasets/pipelines/transforms_3d.py

@@ -1129,11 +1129,13 @@ class BackgroundPointsFilter(object):
         points = input_dict['points']
         gt_bboxes_3d = input_dict['gt_bboxes_3d']
 
-        gt_bboxes_3d_np = gt_bboxes_3d.tensor.numpy()
-        gt_bboxes_3d_np[:, :3] = gt_bboxes_3d.gravity_center.numpy()
+        # avoid groundtruth being modified
+        gt_bboxes_3d_np = gt_bboxes_3d.tensor.clone().numpy()
+        gt_bboxes_3d_np[:, :3] = gt_bboxes_3d.gravity_center.clone().numpy()
+
         enlarged_gt_bboxes_3d = gt_bboxes_3d_np.copy()
         enlarged_gt_bboxes_3d[:, 3:6] += self.bbox_enlarge_range
-        points_numpy = points.tensor.numpy()
+        points_numpy = points.tensor.clone().numpy()
         foreground_masks = box_np_ops.points_in_rbbox(points_numpy,
                                                       gt_bboxes_3d_np)
         enlarge_foreground_masks = box_np_ops.points_in_rbbox(

mmdet3d/models/dense_heads/ssd_3d_head.py

@@ -491,7 +491,7 @@ class SSD3DHead(VoteHead):
             bbox.clone(),
             box_dim=bbox.shape[-1],
             with_yaw=self.bbox_coder.with_rot,
-            origin=(0.5, 0.5, 1.0))
+            origin=(0.5, 0.5, 0.5))
 
         if isinstance(bbox, LiDARInstance3DBoxes):
             box_idx = bbox.points_in_boxes(points)