Self-Similarity Grouping: A Simple Unsupervised Cross Domain Adaptation Approach for Person Re-Identification
https://github.com/SHI-Labs/Self-Similarity-Grouping
youtube 1:28:21
Yang Fu, Yunchao Wei, Guanshuo Wang, Yuqian Zhou, et al.
University of Illinois at Urbana-Champaign, University of Technology Sydney.
关注目标域上特征的相似性。提出了Self-similarity Grouping (SSG)方法，从全局和局部构建多个聚类，进行聚类和训练的迭代过程；在SSG上，基于聚类得到的伪标签提出了半监督的训练策略，实现one-shot domain adaption in an open set setting，将它集成在SSG框架中，进行无监督分支和半监督分支的一起学习，得到SSG++。没有明白具体的训练过程。
Fully Supervised Pretraining
ResNet50 GAP后接两层全连接，FC-2048和FC-#ID，FC-#ID后用交叉熵损失，FC-2048后用三元组损失。
Unsupervised Self-similarity Grouping (SSG)
Clustering-guided Semi-Supervised Training
we employ the unsupervised clustering algorithm on $f_t$ to generate $N_g$ groups. Then, we randomly sample a single image from each group to form a very small sub-dataset $X_g$ with $N_g$ images. Next, we label this small sub-dataset manually and perform labels assignment based on this annotation.
Specifically, we extract features of all images in sub-dataset $X_g$ and obtain three feature vector sets $f_g$ $f_{g-up}$ $f_{g-low}$ and treat each of them as an identity dictionary. Given an unlabeled image $x_t^i$, we find the most similar images from $X_g$ by different cues, whole bodies, upper parts and lower parts, and assign $x_t^i$ with corresponding labels:
Note that we employ the k-reciprocal encoding as the distance metric for similarity measurement.
SSG++
