fcos3d visualization error

[LMerCy]

fcos3d visualization error still exists after updating mmcv to 1.5.2, and i use dev branch of mmdet3d(commit id is e252ff5), according to this issue #1483 (comment)

@filaPro do you have any idea?

Reproduction

python test.py xxxx/configs/fcos3d/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_finetune.py xxx/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_finetune_20210717_095645-8d806dc2.pth --out xxx/res.pkl --show --show-dir xxx/res

I haven't change anything of config and code.

Environment

TorchVision: 0.9.1
OpenCV: 4.5.5
MMCV: 1.5.2
MMCV Compiler: GCC 7.3
MMCV CUDA Compiler: 10.1
MMDetection: 2.24.1
MMSegmentation: 0.24.1
MMDetection3D: 1.0.0rc2+e252ff5
spconv2.0: False

[VVsssssk]

Hi, I think you can refer to this issue #1024.

[LMerCy]

Hi, I think you can refer to this issue #1024.

Thanks，i may underrate the difference between 2d detection and 3d detection. I try to increase score_thr to 0.1 and decrese nms_thr to 0.5, some redundant boxes will be removed which indicates the nms is working(if i contintune to decrease nms_thr to 0.1, most redundant boxes will be removed). It seems due to the depth error, fcos3d needs much smaller nms_thr than 2d detection.