Person Re-Identification Models

written with Pytorch

This is a conclusive repository recording my way of learning pytorch.

Hope it is also helpful for you.

requirement:

  Pytorch version: 1.0.0
  Python version: 3.5.0 and above

Person re-identification is an engineering application of deep neural networks on problems like person search, Multi-person tracking.

For training and testing models, we can configure ./config/config.txt , some parameters are introduced as follows:

model : select a model in path ./model/
loss : input the losses and their weights

It might be upsetting that for training, we have to manually edit ```./loss/init.py''' (line 58-60, 65-67) so as to adjust to different outputs, although this is very easy.

For training and testing , after setting up config.txt:

 $ cd Person-ReID
 $ mkdir log
 $ python3 main.py --cfg config/config.txt

For MGN+RPP or PCB+RPP models, there are 2 steps:

1. set model to mgnrpp/pcbrpp, module to MGN, train 400 epochs.

2. set module to RPP, freeze to 100, load to the checkpoint just saved, train 200 epochs.

---

Test Results

of some models on Market-1501 dataset are listed as the following:

MODEL	dataset	mAP	rank-1	rank-3	rank-5	rank-10
AMG_front	Market1501	0.5887	0.7936	0.8884	0.9178	0.9471
MGN with p2=3&p3=4	Market1501	0.8421	0.9365	0.9688	0.9771	0.9857
ResNet50	Market1501	0.6647	0.8438	0.9138	0.9403	0.9611
ResNet50-mid	Market1501	0.7027	0.8628	0.9267	0.9486	0.9682
PCB	Market1501	0.8128	0.9305	0.9623	0.9724	0.9849
PCB with rollback	Market1501	0.8232	0.9362	0.9659	0.9742	0.984

PS. In this repository, MGN's base model is MobileNetV2, but not the ResNet50.

Rollback is a trick of training by:

Youngmin Ro, Jongwon Choi, Dae Ung Jo, Byeongho Heo, Jongin Lim, Jin Young Choi, " Backbone Can Not be Trained at Once: Rolling Back to Pre-trained Network for Person Re-Identification", CoRR, 2019. (AAAI at 2019 Feb.)