Domain Generalized Person Re-Identification via Cross-Domain Episodic Learning

Abstract

Aiming at recognizing images of the same person across distinct camera views, person re-identification (re-ID) has been among active research topics in computer vision. Most existing re-ID works require collection of a large amount of labeled image data from the scenes of interest. When the data to be recognized are different from the source-domain training ones, a number of domain adaptation approaches have been proposed. Nevertheless, one still needs to collect labeled or unlabelled target-domain data during training. In this paper, we tackle an even more challenging and practical setting, domain generalized (DG) person re-ID. That is, while a number of labeled source-domain datasets are available, we do not have access to any target-domain training data. In order to learn domaininvariant features without knowing the target domain of interest, we present an episodic learning scheme which advances meta learning strategies to exploit the observed source-domain labeled data. The learned features would exhibit sufficient domaininvariant properties while not overfitting the source-domain data or ID labels. Our experiments on four benchmark datasets confirm the superiority of our method over the state-of-the-arts.

Learning to generalize aerial person re‐identification using the meta‐transfer method

“…On another widely studied dataset, DukeMTMC‐reID, the Rank‐1 accuracy has also reached 89.4%, 18 and the mAP has reached 79.6% 17 . This means that there is little room to develop traditional methods, causing the research focus to inevitably shift to more challenging and more practical scenarios 4,5,19,20 …”

“…17 This means that there is little room to develop traditional methods, causing the research focus to inevitably shift to more challenging and more practical scenarios. 4,5,19,20…”

“…Recently, closed‐world Re‐ID 1‐3 has been widely studied and has achieved performance saturation on several datasets. Consequently, the relevant research focus has gradually shifted to the more challenging open‐world Re‐ID problems 4,5 . In various person Re‐ID tasks conducted under closed‐world settings, affected by the development of modern unmanned aerial vehicle (UAV) technology and its unique practical importance, person Re‐ID in aerial imagery 6‐8 has become a research subject worthy of great effort.…”

“…Consequently, the relevant research focus has gradually shifted to the more challenging open-world Re-ID problems. 4,5 In various person Re-ID tasks conducted under closed-world settings, affected by the development of modern unmanned aerial vehicle (UAV) technology and its unique practical importance, person Re-ID in aerial imagery [6][7][8] has become a research subject worthy of great effort. Research on person Re-ID in aerial imagery is not only helpful for criminal investigations, searches for missing persons and the prevention of terror attack, but is also especially useful for conducting person search-and-rescue operations in natural disasters such as earthquakes and fires.…”

Unsupervised generalizable multi-source person re-identification: A Domain-specific adaptive framework

“…Moreover, in poor generalization scenarios, Choi et al [40] explored using batch instance normalization to address overfitting. In addition, Meta-learning [41], [42], [26], [27] is another approach used in network normalization methods. For instance, Song et al [42] develop a weight network for the class-specific classifier, which can map domaininvariant information, and utilize meta-learning to enhance generalization performance on unseen datasets.…”