fast-person-search-baseline

A fast and strong person search baseline!

Prerequisites

• Python 3.9.7
• Pytorch 1.10
• torchvision 0.11.2
• tqdm 4.15.0
• scikit-learn 1.0.1
• pandas 1.4.2

Features

• AMP training
• Large batch size
• Multi-gpu training

Get Started

1. Download this repository by running git clone https://github.com/Niculuse/fast-person-search-baseline.git or clicking the button clone or download.

2. Prepare datasets

   You may download the person search datasets PRW and CUHK-SYSU first, and then prepare the datasets via following commands:

   cd fast-person-search-baseline
   mkdir data

   (1) Market1501 and DukeMTMC

   • Extract the datasets and rename them to prw and sysu, respectively.
   • Copy the folder prw and sysu to the folder data. The data structure should be like:

   data
       prw
           frames/
           annotations/
           ......
       sysu
           frames/
           annotation/
           ...... 

Train the Baseline

You may train a baseline model by following command:

sh train.sh

Before starting multi-gpu training, you should run the script with only one gpu to preprocess the datasets!

With the default settings. You may get results as follows:

PRW:  mAP:48.6%, top-1 86.7%
CUHK-SYSU: mAP:86.7%, top-1 88.4%

You can edit the file train.sh with training settings you want.