songw-zju/HASSC

The official implementation of "Not All Voxels Are Equal: Hardness-Aware Semantic Scene Completion with Self-Distillation" (CVPR 2024)

HASSC

Song Wang, Jiawei Yu, Wentong Li, Wenyu Liu, Xiaolu Liu, Junbo Chen*, Jianke Zhu*

This is the official implementation of Not All Voxels Are Equal: Hardness-Aware Semantic Scene Completion with Self-Distillation (CVPR 2024) [Paper] [Video].

Preparation

SemanticKITTI Download

• The semantic scene completion dataset v1.1 (SemanticKITTI voxel data, 700 MB) from SemanticKITTI website.
• The RGB images (Download odometry data set (color, 65 GB)) from KITTI Odometry website.

Environment Setup

We release the HASSC implementation with VoxFormer-T, please refer the environment setup in the original repo.

Run and Eval

Train the SSC model with our proposed HASSC on 4 GPUs

./tools/dist_train.sh ./projects/configs/hassc/hassc-voxformer-T.py 4

Eval the SSC model with our proposed HASSC on 4 GPUs

./tools/dist_test.sh ./projects/configs/hassc/hassc-voxformer-T.py ./path/to/ckpts.pth 4

Acknowledgement

Many thanks to these excellent open source projects: VoxFormer, mmdetction3d, PointRend

Citations

@inproceedings{wang2024not,
      title={Not All Voxels Are Equal: Hardness-Aware Semantic Scene Completion with Self-Distillation},
      author={Wang, Song and Yu, Jiawei and Li, Wentong and Liu, Wenyu and Liu, Xiaolu and Chen, Junbo and Zhu, Jianke},
      booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
      year={2024}
}