Sequential End-to-end Network for Efficient Person Search

SeqNet

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.

Performance profile:

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

The network structure is simple and suitable as baseline:

net_arch-1

Installation

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

Quick Start

Let's say $ROOT is the root directory.

Download CUHK-SYSU and PRW datasets, and unzip them to $ROOT/data

$ROOT/data
├── CUHK-SYSU
└── PRW

Following the link in the above table, download our pretrained model to anywhere you like, e.g., $ROOT/exp_cuhk
Run an inference demo by specifing the paths of checkpoint and corresponding configuration file. python train.py --cfg $ROOT/exp_cuhk/config.yaml --ckpt $ROOT/exp_cuhk/epoch_19.pth You can checkout the result in demo_imgs directory.

demo

Training

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

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

Note: At present, our script only supports single GPU training, but distributed training will be also supported in 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

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}
}