Skip to content

Latest commit

 

History

History
117 lines (80 loc) · 6.28 KB

README.md

File metadata and controls

117 lines (80 loc) · 6.28 KB
图片名称

Pure SeqNet: PWC PWC

SeqNet with SOLIDER: PWC PWC

This repository hosts the source code of our paper: [AAAI 2021]Sequential End-to-end Network for Efficient Person Search. SeqNet achieves the state-of-the-art performance on two widely used benchmarks and runs at 11.5 FPS on a single GPU. You can find a brief Chinese introduction at zhihu.

SeqNet performance:

Dataset mAP Top-1 Model
CUHK-SYSU 94.8 95.7 model
PRW 47.6 87.6 model

SeqNet with SOLIDER performance:

Dataset mAP Top-1 Model
CUHK-SYSU 95.5 95.8 -
PRW 59.8 86.7 -

The network structure is simple and suitable as baseline:

SeqNet

Updates

[2023/04/10: SOLIDER makes SeqNet better!]: SOLIDER is a Semantic Controllable Self-Supervised Learning Framework to learn general human representations from massive unlabeled human images which can benefit downstream human-centric tasks to the maximum extent. With SOLIDER backbone, SeqNet achieves better results. Please refer to their repo for more details. Nice work!

Installation

Run pip install -r requirements.txt in the root directory of the project.

Quick Start

Let's say $ROOT is the root directory.

  1. Download CUHK-SYSU (google drive or baiduyun) and PRW (google drive or baiduyun) datasets, and unzip them to $ROOT/data
$ROOT/data
├── CUHK-SYSU
└── PRW
  1. Following the link in the above table, download our pretrained model to anywhere you like, e.g., $ROOT/exp_cuhk
  2. Run an inference demo by specifing the paths of checkpoint and corresponding configuration file. python demo.py --cfg $ROOT/exp_cuhk/config.yaml --ckpt $ROOT/exp_cuhk/epoch_19.pth You can checkout the result in demo_imgs directory.

demo.jpg

Training

Pick one configuration file you like in $ROOT/configs, and run with it.

python train.py --cfg configs/cuhk_sysu.yaml

Note:

  • If you are unable to reproduce our results, please check the PyTorch version. Related issues: #26 #29 #31 #32
  • At present, our script only supports single GPU training, but it may support distributed training in the future. By default, the batch size and the learning rate during training are set to 5 and 0.003 respectively, which requires about 28GB of GPU memory. If your GPU cannot provide the required memory, try smaller batch size and learning rate (performance may degrade). Specifically, your setting should follow the Linear Scaling Rule: When the minibatch size is multiplied by k, multiply the learning rate by k. For example:
python train.py --cfg configs/cuhk_sysu.yaml INPUT.BATCH_SIZE_TRAIN 2 SOLVER.BASE_LR 0.0012
  • If your GPU memory is not enough, you can also try strategies such as mixed precision training (FP16), cumulative gradient, and gradient checkpoint. I have tried FP16, which can achieve the same accuracy while saving half of the GPU memory. Unfortunately, I lost this part of the code about FP16. Related PR is welcomed, I can provide some code implementation suggestions.

Tip: If the training process stops unexpectedly, you can resume from the specified checkpoint.

python train.py --cfg configs/cuhk_sysu.yaml --resume --ckpt /path/to/your/checkpoint

Test

Suppose the output directory is $ROOT/exp_cuhk. Test the trained model:

python train.py --cfg $ROOT/exp_cuhk/config.yaml --eval --ckpt $ROOT/exp_cuhk/epoch_19.pth

Test with Context Bipartite Graph Matching algorithm:

python train.py --cfg $ROOT/exp_cuhk/config.yaml --eval --ckpt $ROOT/exp_cuhk/epoch_19.pth EVAL_USE_CBGM True

Test the upper bound of the person search performance by using GT boxes:

python train.py --cfg $ROOT/exp_cuhk/config.yaml --eval --ckpt $ROOT/exp_cuhk/epoch_19.pth EVAL_USE_GT True

Pull Request

Pull request is welcomed! Before submitting a PR, DO NOT forget to run ./dev/linter.sh that provides syntax checking and code style optimation.

Citation

@inproceedings{li2021sequential,
  title={Sequential End-to-end Network for Efficient Person Search},
  author={Li, Zhengjia and Miao, Duoqian},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={35},
  number={3},
  pages={2011--2019},
  year={2021}
}