Structure-Aware Person Search with Self-Attention and Online Instance Aggregation Matching

Installation:

1. Clone the person-PyTorch repository

   git clone https://github.com/gggcy/person_search.git

2. Install - if necessary - the required dependencies:

• Python (tested with 3.7 and 3.6)
• PyTorch (conda install pytorch torchvision -c pytorch - tested with PyTorch 0.4, CUDA 9.0 and 9.2)
• Other python dependencies: numpy, scipy, tensorboard etc.

3. Download datasets. PRW can be downloaded directly. CUHK-SYSU needs to apply to the author's mailbox.

Training the network

1. sh ./setup.sh
2. run dataset/process_prw.py & dataset/process_sysu.py
3. Download the VGG/ResNet50 model pretrained on ImageNet
4. run train.py. This step takes 20 hours to train CUHK-SYSU or 14 hours to train PRW (Single 1080ti)

Testing the network

• run test.py