luhaifeng19947 / SLS_ReID

Sparse Label Smoothing Regularization for semi-supervised Person Re-Identification

SLS ReID

Datasets

• Download Market1501 Dataset
• Download DukeMTMC-reID Dataset
• Download CUHK03 Dataset
• Download VIPeR Dataset

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

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 = {{Ainam}, J.-P. and {Qin}, K. and {Liu}, G. and {Luo}, G.},
    title = "{Sparse Label Smoothing for Semi-supervised Person Re-Identification}",
    journal = {ArXiv e-prints},
    archivePrefix = "arXiv",
    eprint = {1809.04976},
    year = 2018,
    month = sep,
}