[Fix] Avoid loss NaN and fix RandomFlip3D aug in mvxnet (#2462)

* fix rotation in MVXNet

* add comments

* add comments

* update link of mvxnet in inference.md

configs/mvxnet/README.md

@@ -21,8 +21,8 @@ We implement MVX-Net and provide its results and models on KITTI dataset.
 ### KITTI
 
 |                              Backbone                               |  Class  |  Lr schd   | Mem (GB) | Inf time (fps) | mAP  |                                                                                                                                                                                                  Download                                                                                                                                                                                                   |
-| :-----------------------------------------------------------------: | :-----: | :--------: | :------: | :------------: | :---: | :--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------: |
-| [SECFPN](./mvxnet_fpn_dv_second_secfpn_8xb2-80e_kitti-3d-3class.py) | 3 Class | cosine 80e |   6.7    |                | 63.22 | [model](https://download.openmmlab.com/mmdetection3d/v1.0.0_models/mvxnet/dv_mvx-fpn_second_secfpn_adamw_2x8_80e_kitti-3d-3class/dv_mvx-fpn_second_secfpn_adamw_2x8_80e_kitti-3d-3class_20210831_060805-83442923.pth) \| [log](https://download.openmmlab.com/mmdetection3d/v1.0.0_models/mvxnet/dv_mvx-fpn_second_secfpn_adamw_2x8_80e_kitti-3d-3class/dv_mvx-fpn_second_secfpn_adamw_2x8_80e_kitti-3d-3class_20210831_060805.log.json) |
+| :-----------------------------------------------------------------: | :-----: | :--------: | :------: | :------------: | :--: | :---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------: |
+| [SECFPN](./mvxnet_fpn_dv_second_secfpn_8xb2-80e_kitti-3d-3class.py) | 3 Class | cosine 80e |   6.7    |                | 63.5 | [model](https://download.openmmlab.com/mmdetection3d/v1.1.0_models/mvxnet/mvxnet_fpn_dv_second_secfpn_8xb2-80e_kitti-3d-3class/mvxnet_fpn_dv_second_secfpn_8xb2-80e_kitti-3d-3class-8963258a.pth) \| [log](https://download.openmmlab.com/mmdetection3d/v1.1.0_models/mvxnet/mvxnet_fpn_dv_second_secfpn_8xb2-80e_kitti-3d-3class/mvxnet_fpn_dv_second_secfpn_8xb2-80e_kitti-3d-3class-20230424_132228.log) |
 
 ## Citation
 

configs/mvxnet/metafile.yml

@@ -27,5 +27,5 @@ Models:
       - Task: 3D Object Detection
         Dataset: KITTI
         Metrics:
-          mAP: 63.22
-    Weights: https://download.openmmlab.com/mmdetection3d/v1.0.0_models/mvxnet/dv_mvx-fpn_second_secfpn_adamw_2x8_80e_kitti-3d-3class/dv_mvx-fpn_second_secfpn_adamw_2x8_80e_kitti-3d-3class_20210831_060805-83442923.pth
+          mAP: 63.5
+    Weights: https://download.openmmlab.com/mmdetection3d/v1.1.0_models/mvxnet/mvxnet_fpn_dv_second_secfpn_8xb2-80e_kitti-3d-3class/mvxnet_fpn_dv_second_secfpn_8xb2-80e_kitti-3d-3class-8963258a.pth

configs/mvxnet/mvxnet_fpn_dv_second_secfpn_8xb2-80e_kitti-3d-3class.py

@@ -32,6 +32,8 @@ model = dict(
         type='mmdet.FPN',
         in_channels=[256, 512, 1024, 2048],
         out_channels=256,
+        # make the image features more stable numerically to avoid loss nan
+        norm_cfg=dict(type='BN', requires_grad=False),
         num_outs=5),
     pts_voxel_encoder=dict(
         type='DynamicVFE',

docs/en/user_guides/inference.md

@@ -66,7 +66,7 @@ python demo/multi_modality_demo.py ${PCD_FILE} ${IMAGE_FILE} ${ANNOTATION_FILE}
 
 where the `ANNOTATION_FILE` should provide the 3D to 2D projection matrix. The visualization results including a point cloud, an image, predicted 3D bounding boxes and their projection on the image will be saved in `${OUT_DIR}/PCD_NAME`.
 
-Example on KITTI data using [MVX-Net](https://github.com/open-mmlab/mmdetection3d/tree/main/configs/mvxnet) model:
+Example on KITTI data using [MVX-Net model](https://download.openmmlab.com/mmdetection3d/v1.1.0_models/mvxnet/mvxnet_fpn_dv_second_secfpn_8xb2-80e_kitti-3d-3class/mvxnet_fpn_dv_second_secfpn_8xb2-80e_kitti-3d-3class-8963258a.pth):
 
 ```shell
 python demo/multi_modality_demo.py demo/data/kitti/000008.bin demo/data/kitti/000008.png demo/data/kitti/000008.pkl configs/mvxnet/mvxnet_fpn_dv_second_secfpn_8xb2-80e_kitti-3d-3class.py ${CHECKPOINT_FILE} --cam-type CAM2 --show

docs/zh_cn/user_guides/inference.md

@@ -66,7 +66,7 @@ python demo/multi_modality_demo.py ${PCD_FILE} ${IMAGE_FILE} ${ANNOTATION_FILE}
 
 `ANNOTATION_FILE` 需要提供 3D 到 2D 的仿射矩阵，可视化结果会被保存在 `${OUT_DIR}/PCD_NAME`，其中包括点云、图像、预测的 3D 框以及它们在图像上的投影。
 
-在 KITTI 数据上测试 [MVX-Net](https://github.com/open-mmlab/mmdetection3d/tree/main/configs/mvxnet) 模型：
+在 KITTI 数据上测试 [MVX-Net 模型](https://download.openmmlab.com/mmdetection3d/v1.1.0_models/mvxnet/mvxnet_fpn_dv_second_secfpn_8xb2-80e_kitti-3d-3class/mvxnet_fpn_dv_second_secfpn_8xb2-80e_kitti-3d-3class-8963258a.pth)：
 
 ```shell
 python demo/multi_modality_demo.py demo/data/kitti/000008.bin demo/data/kitti/000008.png demo/data/kitti/000008.pkl configs/mvxnet/mvxnet_fpn_dv_second_secfpn_8xb2-80e_kitti-3d-3class.py ${CHECKPOINT_FILE} --cam-type CAM2 --show

mmdet3d/datasets/transforms/transforms_3d.py

@@ -191,7 +191,15 @@ class RandomFlip3D(RandomFlip):
         if 'flip' not in results:
             cur_dir = self._choose_direction()
         else:
+            # `flip_direction` works only when `flip` is True.
+            # For example, in `MultiScaleFlipAug3D`, `flip_direction` is
+            # 'horizontal' but `flip` is False.
+            if results['flip']:
+                assert 'flip_direction' in results, 'flip and flip_direction '
+                'must exist simultaneously'
                 cur_dir = results['flip_direction']
+            else:
+                cur_dir = None
         if cur_dir is None:
             results['flip'] = False
             results['flip_direction'] = None