(Visualization of RenderOcc's prediction, which is supervised only with 2D labels.)

RenderOcc is a novel paradigm for training vision-centric 3D occupancy models only with 2D labels. Specifically, we extract a NeRF-style 3D volume representation from multi-view images, and employ volume rendering techniques to establish 2D renderings, thus enabling direct 3D supervision from 2D semantics and depth labels.

• Installation

• Prepare Dataset

• Train

  # Train RenderOcc with 8 GPUs
  ./tools/dist_train.sh ./configs/renderocc/renderocc-7frame.py 8

• Evaluation

  # Eval RenderOcc with 8 GPUs
  ./tools/dist_test.sh ./configs/renderocc/renderocc-7frame.py ./path/to/ckpts.pth 8

• Visualization

  # Dump predictions
  bash tools/dist_test.sh configs/renderocc/renderocc-7frame.py renderocc-7frame-12e.pth 1 --dump_dir=work_dirs/output
  # Visualization (select scene-id)
  python tools/visualization/visual.py work_dirs/output/scene-xxxx

  (The pkl file needs to be regenerated for visualization.)

• More model weights will be released later.

Many thanks to these excellent open source projects:

• BEVDet, DVGO, Occ3D, SurroundDepth, OpenOccupancy, CVPR2023-Occ-Challenge

Related Projects:

• SurroundOcc, TPVFormer, BEVFormer, VoxFormer, FB-Occ, SimpleOccupancy, OVO

If this work is helpful for your research, please consider citing:

@article{pan2023renderocc,
  title={RenderOcc: Vision-Centric 3D Occupancy Prediction with 2D Rendering Supervision},
  author={Pan, Mingjie and Liu, Jiaming and Zhang, Renrui and Huang, Peixiang and Li, Xiaoqi and Liu, Li and Zhang, Shanghang},
  journal={arXiv preprint arXiv:2309.09502},
  year={2023}
}