Unsupervised Person Re-identification: Clustering and Fine-tuning
https://github.com/hehefan/Unsupervised-Person-Re-identification-Clustering-and-Fine-tuning
Hehe Fan, Liang Zheng, Chenggang Yan, Yi Yang.
Hangzhou Dianzi University, University of Technology Sydney.
相关工作介绍很详细，首次提出在无监督的行人重识别上进行Clustering and Fine-tuning。提出Progressive Unsupervised Learning (PUL)方法，不停地迭代进行clustering和fine-tuning；还进行了样本选择操作，以self-paced learning的思路选择可靠的样本进行训练。
框架
优化目标
下列公式分别表示k-means分配伪标签、选择可靠样本、用伪标签进行分类训练。
算法过程
In practice, we use cosine to measure the similarity between two feature vectors.
To guarantee each cluster contains at least one reliable sample, we choose the nearest feature to the corresponding centroid as the center of the cluster.
When the number of selected reliable samples is saturated, the algorithm will converge.
