Paper link in International Conference on 3D Vision (3DV), 2025
Weijie Ma, Jingwei Jiang, Yang Yang, Zehui Chen, Hao Chen

A novel and effective exploration of geometric modeling with intance representation for the modern LSS-based BEV detection.

This codebase is built upon MMDetection3D and SOLOFusion.

Note: This repository serves as a research reference implementation. Due to the rapid evolution of dependencies and the time elapsed since the original development, the code may require adjustments to work with current environments. We recommend referring to the original MMDetection3D and SOLOFusion documentation for the most up-to-date installation procedures and environment setup.

The main configuration file is located at: configs/lssinst/lssinst.py

Stage-1 Initialization: The configuration automatically loads stage-1 pre-trained weights for initialization.

Usage:

# Training
bash tools/dist_train.sh configs/lssinst/lssinst.py 8

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

@article{lssinst,
  title={LSSInst: Improving Geometric Modeling in LSS-Based BEV Perception with Instance Representation},
  author={Weijie Ma and Jingwei Jiang and Yang Yang and Zehui Chen and Hao Chen},
  journal={International Conference on 3D Vision (3DV)},
  year={2025}
}