3D Scene Understanding Through Local Random Access Sequence Modeling (LRAS)

Installation

See Installation.md for detailed installation instructions.

Checkpoints

You can download the checkpoints using download_checkpoints.sh

Please put the checkpoints/ dir in the repo dir: <lras_3d>/checkpoints/

Usage

This repo provides the example codes for Novel View Synthesis, 3D Object Manipulation, and Depth Estimation.

Novel View Synthesis

python run_nvs.py

3D Object Manipulation

python run_object_motion.py

Depth Estimation

python run_depth.py

Changelog

• First release of the repo with basic NVS, 3D Object Motion, Depth Estimation capability

License

This repository is licensed under the MIT License. See the LICENSE file for details.

Citation

@article{lee20253d,
  title={3D Scene Understanding Through Local Random Access Sequence Modeling},
  author={Lee, Wanhee and Kotar, Klemen and Venkatesh, Rahul Mysore and Watrous, Jared and Chen, Honglin and Aw, Khai Loong and Yamins, Daniel LK},
  journal={arXiv preprint arXiv:2504.03875},
  year={2025}
}