Hi, @Tai-Wang and mmdetection3d team,

This issue continues the discussion in https://github.com/saic-vul/imvoxelnet/issues/7. If you are sure, it would be great to have ImVoxelNet in this codebase, I can make a PR. I would like to clarify some points first.

1. SUN RGB-D. Unfortunately there are 2 monocular benchmarks: PerspectiveNet with 30 classes, Total3DUnderstanding with 37 classes + camera pose estimation + 3d room layout estimation. In our paper we also present metrics for your VoteNet benchmark with 10 classes. Do you think this VoteNet option is enough for the beginning? One more question here. For SUN RGB-D we use differentiable rotated IoU from here. Do you want to adapt in to mmdetection3d?

2. ScanNet. Current ScanNetDataset doesn't deal with RGB images. Do we need separate MultiViewScanNet dataset and MultiViewPipeline handling random images sampling on training? Also dataset preparation instruction need to be updated.

3. KITTI. No questions here.

4. nuScenes. Following OFTNet and MonoDIS we compare only on car category. I think it not enough interesting for community and we don't need to make separate benchmark here.