Giters

SamsungLabs / imvoxelnet

[WACV2022] ImVoxelNet: Image to Voxels Projection for Monocular and Multi-View General-Purpose 3D Object Detection

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Support Three Classes with Kitti Dataset and Be Compatible with Official MMDetection3D

haibao-yu opened this issue · comments

commented

Describe the feature

We have released a dataset called DAIR-V2X, and we follow your config and mmdetection3d to implement many imvoxelnet benchmarks on our dataset. More details about the dataset can refer to the paper "DAIR-V2X: A Large-Scale Dataset for Vehicle-Infrastructure Cooperative 3D Object Detection", which is accepted by CVPR2022. More details about our benchmark config can refer to https://github.com/AIR-THU/DAIR-V2X/blob/main/configs/sv3d-inf/imvoxelnet/trainval_config.py.

Now, we are trying to make the config compatible with the official mmdetection3d, and make the config support three classes with Kitti format. But there is large performance fluctuation. So, have you had any related try?

commented

Hi @haibao-yu ,

We tried 3 classes on KITTI at some point (just a very few training runs), but the performance was very low. Mainly because of small sizes of BEV projection of other 2 classes. Btw, official mmdetection3d supports 1 class benchmark.

commented

Hi @haibao-yu ,

We tried 3 classes on KITTI at some point (just a very few training runs), but the performance was very low. Mainly because of small sizes of BEV projection of other 2 classes. Btw, official mmdetection3d supports 1 class benchmark.

Thanks for your reply. We have collaborated with mmdetection3d to get the proper 3class config. The config is available at https://github.com/AIR-THU/DAIR-V2X/blob/main/configs/sv3d-inf/imvoxelnet/trainval_config.py now.