Person Search

✨ News: We release the source code of the current state-of-the-art model SeqNet(AAAI 2021), which achieves 🏆 94.8% mAP on CUHK-SYSU.

Introduction

A pytorch implementation for CVPR 2017 "Joint Detection and Identification Feature Learning for Person Search".

The code is based on the offcial caffe version.

You can find a better one achieving about 85% mAP in mmdetection branch!

Note: The implementaion of Faster R-CNN in mmdetection branch is better than that described in original paper.

Highlights

• Simpler code: After reduction and refactoring, the current version is simpler and easier to understand.
• Pure Pytorch code: Numpy is not used, except for data loading.

Installation

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

torchvision must be greater than 0.3.0, as we need torchvision.ops.nms

Quick Start

Let's say $ROOT is the root directory.

1. Download CUHK-SYSU (google drive or baiduyun) dataset, unzip to $ROOT/data/dataset/
2. Download our trained model (google drive or baiduyun) (extraction code: uuti) to $ROOT/data/trained_model/

After the above two steps, the directory structure should look like this:

$ROOT/data
├── dataset
│   ├── annotation
│   ├── Image
│   └── README.txt
└── trained_model
    └── checkpoint_step_50000.pth

BTW, $ROOT/data saves all experimental data, include: dataset, pretrained model, trained model, and so on.

3. Run python tools/demo.py --gpu 0 --checkpoint data/trained_model/checkpoint_step_50000.pth. And then you can checkout the result in imgs directory.

Train

1. Prepare dataset as described in Quick Start section.
2. Download pretrained model (google drive or baiduyun) (extraction code ucnw) to $ROOT/data/pretrained_model/
3. python tools/train_net.py --gpu 0
4. Trained model will be saved to $ROOT/data/trained_model/

You can check the usage of train_net.py by running python tools/train_net.py -h

Test

python tools/test_net.py --gpu 0 --checkpoint data/trained_model/checkpoint_step_50000.pth

The result should be around:

Search ranking:
   mAP = 76.78%
   Top- 1 = 77.48%
   Top- 5 = 88.48%
   Top-10 = 91.52%