youngminPIL / Rollback

Backbone Can Not be Trained at Once: Rolling Back to Pre-trained Network for Person Re-Identification

Official Pytorch implementation of paper:

Backbone Can Not be Trained at Once: Rolling Back to Pre-trained Network for Person Re-Identification(AAAI 2019).

Project page

Environment

Python 3.6, Pytorch 0.4.1, Torchvision, tensorboard(optional)

Train

Default setting:

• Architecture: ResNet-50
• Dataset: Market-1501
• Batch size: 32
• Image size: 288X144
• Train 4 period.

prepare

The dataset path should be changed to your own path.

Market-1501 dataset are available on http://www.liangzheng.org/Project/project_reid.html

prepare.py 

train network on the each periods.

Train model in period 1. This is a baseline of our algorithm.

The dataset path(data_dir='/home/ro/Reid/Market/pytorch') should be changed to your own path.

train_resnet_p1.py

Each period should be trained on the results of previous training.

train_resnet_p2.py

train_resnet_p3.py

train_resnet_p4.py

Test

The test will be done when you complete your trainung up to period 4.

The dataset path(test_dir='/home/ro/Reid/Market/pytorch') should be changed to your own path.

test_resnet.py

Citation

@inproceedings{rollback_v1,
	title = {Backbone Can Not be Trained at Once: Rolling Back to Pre-trained Network for Person Re-Identification
},
	author = {Youngmin Ro, Jongwon Choi, Dae Ung Jo, Byeongho Heo, Jongin Lim, Jin Young Choi},
	booktitle = {AAAI},
	year = {2019}
}

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.)