This repository is an official implementation of Far3D.
Expanding existing methods directly to cover long distances poses challenges such as heavy computation costs and unstable convergence. To address these limitations, we proposes a novel sparse query-based framework, dubbed Far3D. By utilizing high-quality 2D object priors, we generate 3D adaptive queries that complement the 3D global queries. To efficiently capture discriminative features across different views and scales for long-range objects, we introduce a perspective-aware aggregation module. Additionally, we propose a range-modulated 3D denoising approach to address query error propagation and mitigate convergence issues in long-range tasks.
Our pipeline follows StreamPETR, and you can follow Get Started step by step.
Quick Train & Evaluation
Train the model
tools/dist_train.sh projects/configs/far3d.py 8 --work-dir work_dirs/far3d/
Evaluation
tools/dist_test.sh projects/configs/far3d.py work_dirs/far3d/iter_82548.pth 8 --eval bbox
Model | Backbone | Input size | mAP | CDS | Config | Download |
---|---|---|---|---|---|---|
BEVStereo | VoV-99 | (960, 640) | 0.146 | 0.104 | -- | -- |
SOLOFusion | VoV-99 | (960, 640) | 0.149 | 0.106 | -- | -- |
PETR | VoV-99 | (960, 640) | 0.176 | 0.122 | -- | -- |
Sparse4Dv2 | VoV-99 | (960, 640) | 0.189 | 0.134 | -- | -- |
StreamPETR | VoV-99 | (960, 640) | 0.203 | 0.146 | -- | -- |
Far3D | VoV-99 | (960, 640) | 0.244 | 0.181 | config | model/log |
Notes
We thank these great works and open-source codebases:
If you find Far3D is useful in your research or applications, please consider giving us a star 🌟 and citing it by the following BibTeX entry.
@article{jiang2023far3d,
title={Far3D: Expanding the Horizon for Surround-view 3D Object Detection},
author={Jiang, Xiaohui and Li, Shuailin and Liu, Yingfei and Wang, Shihao and Jia, Fan and Wang, Tiancai and Han, Lijin and Zhang, Xiangyu},
journal={arXiv preprint arXiv:2308.09616},
year={2023}
}