Skip to content

Repo for BMVC2020 paper "BCaR: Beginner Classifier as Regularization Towards Generalizable Re-ID"

License

Notifications You must be signed in to change notification settings

hitachi-rd-cv/bcar

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

BCaR: Beginner Classifier as Regularization Towards Generalizable Re-ID

Official repository of our BMVC2020 paper. The code is implemented by PyTorch and some other libraries.

overview

Requirements

Experiments were conducted on Ubuntu 18.04 with Python 3.6 and PyTorch v1.3.1. You can install the dependencies by the following command:

pip install -r requirements.txt

Training and test

1. Data preparation

In our experiments, the following datasets are used for the training and test:

  • CUHK02
  • CUHK03
  • Market1501
  • DukeMTMC-reID
  • PersonSearch
  • VIPeR
  • PRID
  • GRID
  • i-LIDS

Download and extract all the datasets and place them to data/original. For training, the data have to be relocated. Using the following command, the data are automatically relocated:

./scripts/relocate_dataset.sh

Note that some of the datasets might be removed from the Internet. We cannot provide them so please ask the dataset authors about the status.

2. Start training and test

We have prepared the training scripts for each of the experiments described in our paper. You can use them if the data are relocated by our script. If not, change the train_targets and test_targets options. Our scripts can be used as follows:

Domain generalization setting with the MobileNetV2 backbone

./scripts/train_mobilenet_v2.sh

Domain generalization setting with the ResNet50 backbone

./scripts/train_resnet50.sh

Supervised setting for Market1501

./scripts/train_market.sh

Supervised setting for DukeMTMC-reID

./scripts/train_duke.sh

The tests are carried out at the end of each epoch.

3. Check results

Training and test results are saved to the directory specified at the start of training with the log_path option. You can check the results with Tensorboard. For example, if the option is set to logs_mobilenet_v2, you can start Tensorboard with the following command:

tensorboard --logdir logs_mobilenet_v2

Results

Domain generalization setting

For each dataset, the rank-1 accuracy is shown.

MobileNetV2 backbone

VIPeR PRID GRID i-LIDS
DIMN 51.2 39.2 29.3 70.2
DualNorm 53.9 60.4 41.4 74.8
DualNorm + BCaR 57.3 62.0 42.3 80.0

ResNet50 backbone

VIPeR PRID GRID i-LIDS
DualNorm 59.4 69.6 43.7 78.2
DualNorm + BCaR 65.8 70.2 52.8 81.3

Supervised setting

Market1501

Rank-1 Rank-5 Rank-10 mAP
DualNorm 82.6 91.7 95.3 57.2
DualNorm + BCaR 87.6 95.5 97.1 65.7

DukeMTMC-reID

Rank-1 Rank-5 Rank-10 mAP
DualNorm 71.2 82.5 86.3 48.3
DualNorm + BCaR 78.6 86.5 89.6 55.2

Citation

Please consider citing our paper if it helps your research:

@inproceedings{masato_bmvc2020,
author = {Tamura, Masato and Yoshinaga, Tomoaki},
title = {BCaR: Beginner Classifier as Regularization Towards Generalizable Re-ID},
booktitle={BMVC},
year = {2020},
month = {September}
}

Acknowledgement

Some of the codes are build upon DualNorm. Thanks them for the great work!

Releases

No releases published

Packages

No packages published