README.md

June 7, 2026 · View on GitHub

LegoOcc: Monocular Open Vocabulary Occupancy Prediction for Indoor Scenes

CVPR 2026 Oral

Changqing Zhou1, Yueru Luo2, Han Zhang1, Zeyu Jiang1, Changhao Chen1 ✉

1The Hong Kong University of Science and Technology (Guangzhou)
2The Chinese University of Hong Kong, Shenzhen

✉ Corresponding author.

Project Page | Paper

LegoOcc tackles monocular open-vocabulary 3D semantic occupancy prediction in large-scale indoor scenes under geometry-only supervision. It represents scenes as Language-Embedded Gaussians, introduces an opacity-aware Poisson Gaussian-to-Occupancy operator for stable volumetric aggregation, and adopts Progressive Temperature Decay to strengthen Gaussian-language alignment during training.

The framework is designed for open-vocabulary indoor occupancy reasoning from monocular observations, bridging sparse Gaussian scene modeling and language-aware 3D occupancy prediction.

News

  • [2026.05] :rocket: Code is released
  • [2026.04] :microphone: LegoOcc was accepted to CVPR 2026 (Oral).

Documentation

For setup and usage details, please refer to the documents under docs/:

  • docs/install.md: environment setup, dependency installation, and pretrained component preparation.
  • docs/data.md: dataset preparation for OccScanNet, folder structure, and symbolic link setup.
  • docs/train_eval.md: training workflow, configuration notes, and runtime environment variables.

Getting Started

  1. Follow docs/install.md to prepare the environment.
  2. Follow docs/data.md to organize the dataset.
  3. Follow docs/train_eval.md to launch training.

Citation

If you find this work useful, please consider citing:

@inproceedings{zhou2026monocular,
  title={Monocular open vocabulary occupancy prediction for indoor scenes},
  author={Zhou, Changqing and Luo, Yueru and Zhang, Han and Jiang, Zeyu and Chen, Changhao},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={21627--21637},
  year={2026}
}

We recommend checking out the following related projects: