PyTorch Implementation of OIMNet++ (ECCV 2022)

This is an official PyTorch implementation of "OIMNet++: Prototypical Normalization and Localization-aware Learning for Person Search", ECCV 2022.

For more details, visit our project site or see our paper.

Requirements

• Python 3.8
• PyTorch 1.7.1
• GPU memory >= 22GB

Features

• Re-implementation of vanilla OIMNet
• Using AMP to train with larger batch size with limited GPU memory

Getting Started

First, clone our git repository.

Docker

We highly recommend using our Dockerfile to set up the environment.

# build docker image
$ docker build -t oimnetplus:latest . 

# execute docker container
$ docker run --ipc=host -it -v <working_dir>:/workspace/work -v <dataset_dir>:/workspace/dataset -w /workspace/work oimnetplus:latest /bin/bash 

Prepare datasets

Download PRW and CUHK-SYSU datasets.
Modify the dataset directories below if necessary.

• PRW: L4 of configs/prw.yaml
  
• CUHK-SYSU: L3 of configs/ssm.yaml
  

Your directories should look like:

    <working_dir>
    OIMNetPlus
    ├── configs/
    ├── datasets/
    ├── engines/
    ├── losses/
    ├── models/
    ├── utils/
    ├── defaults.py
    ├── Dockerfile
    └── train.py
    
    <dataset_dir>
    ├── CUHK-SYSU/
    │   ├── annotation/
    │   ├── Image/
    │   └── ...
    └── PRW-v16.04.20/
        ├── annotations/
        ├── frames/
        ├── query_box/
        └── ...

Training and Evaluation

• OIMNet++
  $ python train.py --cfg configs/prw.yaml
$ python train.py --cfg configs/ssm.yaml

• OIMNet+++
  $ python train.py --cfg configs/prw.yaml MODEL.ROI_HEAD.AUGMENT True
$ python train.py --cfg configs/ssm.yaml MODEL.ROI_HEAD.AUGMENT True

• OIMNet
  $ python train.py --cfg configs/prw.yaml MODEL.ROI_HEAD.NORM_TYPE 'none' MODEL.LOSS.TYPE 'OIM'
$ python train.py --cfg configs/ssm.yaml MODEL.ROI_HEAD.NORM_TYPE 'none' MODEL.LOSS.TYPE 'OIM'

By running the commands, evaluation results and training losses will be logged into a .txt file in the output directory.

Credits

Our person search implementation is heavily based on Di Chen's NAE and Zhengjia Li's SeqNet.
ProtoNorm implementation is based on ptrblck's manual BatchNorm implementation here.