FastReID is a research platform that implements state-of-the-art re-identification algorithms. It is a groud-up rewrite of the previous verson, reid strong baseline.
- [Aug 2020] Model Distillation is supported, thanks for guan'an wang's contribution.
- [Aug 2020] ONNX/TensorRT converter is supported.
- [Jul 2020] Distributed training with multiple GPUs, it trains much faster.
- [Jul 2020]
MAX_ITERin config meansepoch, it will auto scale to maximum iterations. - Includes more features such as circle loss, abundant visualization methods and evaluation metrics, SoTA results on conventional, cross-domain, partial and vehicle re-id, testing on multi-datasets simultaneously, etc.
- Can be used as a library to support different projects on top of it. We'll open source more research projects in this way.
- Remove ignite(a high-level library) dependency and powered by PyTorch.
We write a chinese blog about this toolbox.
See INSTALL.md.
The designed architecture follows this guide PyTorch-Project-Template, you can check each folder's purpose by yourself.
See GETTING_STARTED.md.
Learn more at out documentation. And see projects/ for some projects that are build on top of fastreid.
We provide a large set of baseline results and trained models available for download in the Fastreid Model Zoo.
We provide some examples and scripts to convert fastreid model to Caffe, ONNX and TensorRT format in Fastreid deploy.
Fastreid is released under the Apache 2.0 license.
If you use Fastreid in your research or wish to refer to the baseline results published in the Model Zoo, please use the following BibTeX entry.
@article{he2020fastreid,
title={FastReID: A Pytorch Toolbox for General Instance Re-identification},
author={He, Lingxiao and Liao, Xingyu and Liu, Wu and Liu, Xinchen and Cheng, Peng and Mei, Tao},
journal={arXiv preprint arXiv:2006.02631},
year={2020}
}