Exploiting Robust Memory Features for Unsupervised Reidentification

"2022/05/10" the conference paper code.

Model

• Our architecture consists of three modules: the backbone feature module, the cluster generation pseudo-labeling module, and cluster memory module.

• Innovation of Our Method.

Train

• The batchsize of 64 is the single-GPU training, and the 256 size for multi-GPU training.

CUDA_VISIBLE_DEVICES=0 python usl.py -b 64 -a resnet50 -d veri --iters 400 --momentum 0.1 --eps 0.7 --num-instances 16 --height 224 --width 224 --use-hard 
CUDA_VISIBLE_DEVICES=0 python usl.py -b 64 -a resnet50 -d msmt17 --iters 400 --momentum 0.1 --eps 0.7 --num-instances 16 --use-hard 
    
CUDA_VISIBLE_DEVICES=0,1,2,3 python usl.py -b 256 -a resnet50 -d veri --iters 400 --momentum 0.1 --eps 0.7 --num-instances 16 --height 224 --width 224 --use-hard 
CUDA_VISIBLE_DEVICES=0,1,2,3 python usl.py -b 256 -a resnet50 -d msmt17 --iters 400 --momentum 0.1 --eps 0.7 --num-instances 16 --use-hard 

Results

• ResNet50 was used as our backbone network and the imagenet pre-trained model was loaded.
• VeRi776.

• MSMT17.

Link

• our paper Exploiting Robust Memory Features for Unsupervised Reidentification

references

@InProceedings{10.1007/978-3-031-18910-4_52,
author="Lian, Jiawei
and Wang, Da-Han
and Du, Xia
and Wu, Yun
and Zhu, Shunzhi",
title="Exploiting Robust Memory Features for Unsupervised Reidentification",
booktitle="Pattern Recognition and Computer Vision",
year="2022",
publisher="Springer Nature Switzerland",
address="Cham",
pages="655--667",
isbn="978-3-031-18910-4"
}

Acknowledgements

---

Acknowledgement to the following open source projects, and we have listed them.

• deepcluster
• cluster-contrast-reid