Skip to content
Sparse Label Smoothing Regularization for Person Re-Identification
Python
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
.idea For generating GAN labels (gan%.list) - call generate_labels_for_gan May 13, 2019
DCGAN Adding results to readme Jul 18, 2018
clustering
images
README.md For generating GAN labels (gan%.list) - call generate_labels_for_gan May 13, 2019
create_dataset.py Fresh install Jul 8, 2018
cuhk03_rerank.py Fresh install Jul 8, 2018
data_generator.py
data_manager.py Add results to readme Jul 17, 2018
dataset.py For generating GAN labels (gan%.list) - call generate_labels_for_gan May 13, 2019
dataset_loader.py
eval_cuhk03.py Add results to readme Jul 17, 2018
eval_viper.py Fresh install Jul 8, 2018
evaluate.py Fresh install Jul 8, 2018
evaluate_gpu.py
evaluate_rerank.py
model.py Adding results to readme Jul 18, 2018
prepare.py
prepare_gan_data.py For generating GAN labels (gan%.list) - call generate_labels_for_gan May 13, 2019
random_erasing.py
re_index.py Adding results to readme Jul 18, 2018
re_ranking.py
sls_train.py
test.py
test_cuhk03.py
test_viper.py Fresh install Jul 8, 2018
train.py
utils.py Fresh install Jul 8, 2018

README.md

SLS ReID

Datasets

Dataset preperation

Change the download_path to point to your dataset folder. Comment out multi-query except for Market1501 dataset.

python prepare.py
python re_index.py

Training

  • Train the baseline
python train.py 
  • Train SLS_ReID
python sls_train.py 

Add --use_dense argument to train using DenseNet121 architecture

Pre-trained models

Dataset Dense Baseline ResNet Baseline Dense SLS_ReID ResNet SLS_ReID
Market-1501 market/dense.pth market/resnet.pth cuhk03/dense_slsreid.pth market/resnet_slsreid.pth
CUHK03 cuhk03/dense.pth cuhk03/resnet.pth cuhk03/dense_slsreid.pth cuhk03/resnet_slsreid.pth
VIPeR viper/dense.pth viper/resnet.pth viper/dense_slsreid.pth viper/resnet_slsreid.pth
DukeMTMCReID duke/dense.pth duke/resnet.pth duke/dense_slsreid.pth duke/resnet_slsreid.pth

To generate the GAN label (gan%.list ie gan0.list, gan1.list and gan2.list for three cluster ), run generate_labels_for_gan:

python prepare_gan_data.py

Testing

python test_cuhk03.py --model_path ./cuhk03/model.pth --use_dense
python eval_cuhk03.py
----------
python test_viper.py --model_path ./viper/model.pth --use_dense
%Add --re_rank to get re-ranking with k-reciprocal encoding
----------
python test.py --model_path ./market/model.pth --use_dense
python evaluate.py
python evaluate_rerank.py 
%Add --multi for multi-query evaluation

Currents results

Dataset Rank 1 Rank 5 Rank 10 Rank 20 mAP
CUHK03-Dense Baseline 67.92% 90.94% 95.35% 97.78% 78.10%
CUHK03-Dense SLS_ReID 84.32% 97.13% 98.92% 99.63% 89.92%
CUHK03-ResNet Baseline 75.02% 95.08% 97.92% 99.11% 83.87%
CUHK03-ResNet SLS_ReID 90.99% 98.24% 99.25% 99.74% 94.18%
VIPeR-Dense Baseline 63.45% 72.78% 79.11% 86.23% -
VIPeR-Dense SLS_ReID 67.41% 81.01% 88.61% 93.51% -
Market-1501-Dense Baseline 90.05% 96.82% 98.10% 98.81% 74.16%
Market-1501-Dense SLS_ReID 92.43% 97.27% 98.40% - 79.08%
Market-1501-ResNet Baseline 87.29% 95.57% 96.94% - 69.70%
Market-1501-ResNet SLS_ReID 89.16% 95.78% 97.33% - 75.15%
DukeMTMC-ReID-Dense Baseline 79.67% 89.85% 92.86% 95.11% 63.19%
DukeMTMC-ReID-Dense SLS_ReID 82.94% 91.69% 94.43% 95.96% 67.78%
DukeMTMC-ReID-ResNet Baseline 76.66% 87.83% 91.47% 93.76% 58.35%
DukeMTMC-ReID-ResNet SLS_ReID 76.53% 88.15% 91.02% 93.54% 60.79%

Multi-query results for Market-1501 dataset

Dataset Rank 1 Rank 5 Rank 10 mAP
ResNet baseline 91.27% 96.85% 98.19% 76.94%
DenseNet baseline 92.90% 97.89% 98.69% 81.22%
Resnet SLS_ReID 92.25% 97.51% 98.34% 81.92%
Dense SLS_ReID 94.06% 98.16% 98.84% 85.20%

References

@ARTICLE{AinamSLSR2018,
author={J. {Ainam} and K. {Qin} and G. {Liu} and G. {Luo}}, 
journal={IEEE Access}, 
title={Sparse Label Smoothing Regularization for Person Re-Identification}, 
year={2019}, 
volume={7},  
pages={27899-27910},  
doi={10.1109/ACCESS.2019.2901599}, 
ISSN={2169-3536}
}
You can’t perform that action at this time.