Joint Disentangling and Adaptation for Cross-Domain Person Re-Identification https://arxiv.org/pdf/2007.10315.pdf 源码：https://github.com/NVlabs/DG-Net-PP
Solving cross-domain problems also belongs to image migration. The features are divided into identity-related information and irrelevant information, which are actually structural features and appearance color features, and then use source domain tags as supervision information to train on the target domain.
Identity-Guided Human Semantic Parsing Learning for Person Re-Identification https://arxiv.org/pdf/2007.13467.pdf
Supervised learning, by aligning body parts and adding the attributes of the backpack to the comparison, the body part alignment is based on adaptability, using the network to determine different parts of the body, and then using feature vectors to express them.
Multiple Expert Brainstorming for Domain Adaptive Person Re-identification https://arxiv.org/pdf/2007.01546.pdf
Use multiple models to train on the source domain, and then use the average features of multiple models as features on the target domain to perform clustering. The clustering process is to cluster on the entire target set. How many times to cluster is set in the early stage of.
Global Distance-distributions Separation for Unsupervised Person Re-identification https://arxiv.org/pdf/2006.00752v1.pdf
It is to optimize the network at the level of the loss function. Compared with triplet loss, this loss is to bring the same closer and different ones from a global perspective. The compared features are the features extracted on the target data set.
Interpretable and Generalizable Person Re-identification with Query-adaptive Convolution and Temporal Lifting https://arxiv.org/pdf/1904.10424v3.pdf
The previous algorithm did not consider the relationship between the two input pictures to be matched. This algorithm subdivides the final feature map and performs a small range of matching one by one. A query must match all gallery pictures. Comparison of improvements in metrics.
Unsupervised Domain Adaptation with Noise Resistible Mutual-Training for Person Re-identification https://zhaoj9014.github.io/pub/1391.pdf
From the perspective of improving the accuracy of pseudo-labels in the target set, a stable pseudo-label can be obtained through mutual learning through the output of multiple networks. When screening accurate tags, select credible tags based on the triplet distance.
Rethinking the Distribution Gap of Person Re-identification with Camera-based Batch Normalization https://arxiv.org/pdf/2001.08680.pdf 源码：https://github.com/automan000/Camera-based-Person-ReID
The statistical information in the same camera is counted, and then it is ensured that all the statistical information is standardized into one form, so that the images of the same camera are processed in the same way.
Joint Visual and Temporal Consistency for Unsupervised Domain Adaptive Person Re-Identification https://arxiv.org/pdf/2007.10854.pdf
On the target data set, in a mini batch, treat different pictures as one category, multiply the features with the classifier to generate scores, the classifier is similar to the feature bank, and this picture is data-enhanced, one picture Several different pictures were generated as the objects of comparison. Globally, it is still based on clustering, but this time clustering is not only based on the similarity of features, but also on empty information.