https://github.com/sjtuzq/person_search_gcn

[CVPR2019, Oral] Learning Context Graph for Person Search
https://github.com/sjtuzq/person_search_gcn

graph-convolutional-networks person-search re-id

Last synced: 14 days ago
JSON representation

[CVPR2019, Oral] Learning Context Graph for Person Search

• Host: GitHub
• URL: https://github.com/sjtuzq/person_search_gcn
• Owner: sjtuzq
• Created: 2019-03-18T14:39:58.000Z (over 5 years ago)
• Default Branch: master
• Last Pushed: 2019-12-04T05:10:50.000Z (almost 5 years ago)
• Last Synced: 2024-08-02T08:10:14.860Z (4 months ago)
• Topics: graph-convolutional-networks, person-search, re-id
• Language: Python
• Homepage:
• Size: 38.1 KB
• Stars: 129
• Watchers: 5
• Forks: 19
• Open Issues: 6
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• awesome-gcn - sjtuzq/person_search_gcn

README

        
# Person Search GCN Project

This project implements Context Graph of paper [Learning Context Graph for Person Search](https://arxiv.org/abs/1904.01830) (CVPR 2019 Oral). 

This repository shows how we can use graph convolution model to employ context information and improve person search performance. 

## Environment:

python(3.6),

pytorch(0.4.1),

numpy(1.14.1), 

matplotlib(3.0.2),

tqdm,

pickle

## Preparation

1. Clone this repo 

  ```Shell

  git clone https://github.com/sjtuzq/person_search_gcn.git

  ```

2. modify data_path, log_path and neighbor_num in config.py

3. download dataset [here](https://drive.google.com/open?id=1-pjZd-bZFTqV2F_34jr0q77-iEmjE4P5), then put it into data_path folder. The features are generated by the method in [Joint Detection and Identification Feature Learning for Person Search](https://arxiv.org/abs/1604.01850), whose code can be found [here](https://github.com/ShuangLI59/person\_search).

4. prepare dataset 

  generate the raw data feature into the paired form: persons in the same images are grouped together.

  ```Shell

  python prepare.py

  ```

## Experiments

1. train and test the gcn model

  train graph convolution model: with the pair selected by distance.

  ```Shell

  python train_gcn1.py

  ```

  The top-1 acc can be imporved from 78.6% to 81.6%, mAP from 75.4% to 80.2%.