SuperOcc: Toward Cohesive Temporal Modeling for Superquadric-based Occupancy Prediction

This repository contains the weights for SuperOcc, a novel framework for superquadric-based 3D occupancy prediction.

Paper | Code

Method

SuperOcc addresses challenges in 3D occupancy prediction for autonomous driving by leveraging 3D superquadric representations. It incorporates three key designs:

1. Cohesive Temporal Modeling: Simultaneously exploits view-centric and object-centric temporal cues.
2. Multi-superquadric Decoding: Enhances geometric expressiveness without sacrificing query sparsity.
3. Efficient Splatting: Improves computational efficiency for superquadric-to-voxel conversion.

Benchmark Results

Results on Occ3D:

Models	Epochs	Q	mIoU	RayIoU	FPS
SuperOcc-T	48	600	36.1	42.5	30.3
SuperOcc-S	48	1200	36.9	43.0	28.2
SuperOcc-M	48	2400	37.4	43.6	18.8
SuperOcc-L	48	3600	38.1	44.0	12.7

Results on SurroundOcc:

Models	Epochs	Q	IoU	mIoU	FPS
SuperOcc-T	24	600	34.91	22.48	30.3
SuperOcc-S	24	1200	35.63	23.12	28.2
SuperOcc-M	24	2400	36.99	23.95	18.8
SuperOcc-L	24	3600	38.13	24.55	12.7

Citation

If you find this work helpful for your research, please consider citing the following BibTeX entry.

@article{yu2026superocc,
  title={SuperOcc: Toward Cohesive Temporal Modeling for Superquadric-based Occupancy Prediction},
  author={Yu, Zichen and Liu, Quanli and Wang, Wei and Zhang, Liyong and Zhao, Xiaoguang},
  journal={arXiv preprint arXiv:2601.15644},
  year={2026}
}