参考链接:Introduction to Mutual Mean-Teaching (ICLR 2020), Structured Domain Adaptation, VisDA-2020 Solution (ECCVW 2020), Self-paced Contrastive Learning (NeurIPS 2020), OpenUnReID Codebase.
领域自适应(Domain Adaptive)
Common scenarios: City A > City B Synthetic > Real-world 在有标签的源域(labeled source domain)与无标签的目标域(unlabeled target domain)进行训练,目的是在目标域取得较好的性能
无监督学习(Unsupervised learning)
No source domain and Unlabeled targrt domain 无监督行人重识别是从预训练的Imagenet网络开始,可以理解为从imagenet的domain向reid的domain迁移 无监督预训练任务是从随机初始化的网络开始
Object Re-identification(Re-ID)
Probe object Gallery images captured from multiple cameras 对每张图像进行特征提取,对每张probe image 与 gallery images计算距离(distance computation),然后进行相似度排序(feature ranking) 最重要的是如何为图像提取具有辨别性的特征(learn discriminative features in varying conditions.) 不同光照条件,角度
|