yxgeee / sda

Structured Domain Adaptation with Online Relation Regularization for Unsupervised Person Re-ID

Structured Domain Adaptation (SDA)

[Paper] [Project]

This repository contains the implementation of Structured Domain Adaptation with Online Relation Regularization for Unsupervised Person Re-ID

We are still sorting out the codes of SDA, which will be released later.

Citation

If you find this work useful for your research, please cite our paper

@misc{ge2020structured,
    title={Structured Domain Adaptation with Online Relation Regularization for Unsupervised Person Re-ID},
    author={Yixiao Ge and Feng Zhu and Rui Zhao and Hongsheng Li},
    year={2020},
    eprint={2003.06650},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

sda's Issues

您好

hello, when will the new code be released ?

GAN implementation details

Hello,

I'm working on implementing this paper. I've taken the code from your MMT-plus repository, and refitted it such that the training algorithm fits Alg(1) from the SDA paper (translating batched on the fly during joint training). Also, I'm using a single mean teacher for the encoder updates with kmeans instead of MMT (just trying to recreate results the original SDA paper results). In the paper the details for the implementation are pretty clear, but the cycle GAN architecture isn't really mentioned. Could you give details as to which GAN style you're using (LSGAN, Wasserstein, spectral, etc), which architectures you're using for generators, discriminators, if the learning rate decay is applied to both the cycleGAN and the target encoder. If you want, we could call and discuss. Let me know.

RC training loss.

Hello,

I'm implementing the original paper, and I've adapted the code from the MMT-plus repository. I've notice that during training, the rc loss converges very quickly to 0.7 and doesn't change. Is this normal from your experience training with SDA? I've included a picture with some other losses for scale.

hello, when will the code be released ?

Source code

when will the source code be released?