Reading Relevant Feature from Global Representation Memory for Visual Object Tracking

arXiv:2402.14392v1 Announce Type: new
Abstract: Reference features from a template or historical frames are crucial for visual object tracking. Prior works utilize all features from a fixed template or memory for visual object tracking. However, due to the dynamic nature of videos, the required reference historical information for different search regions at different time steps is also inconsistent. Therefore, using all features in the template and memory can lead to redundancy and impair tracking performance. To alleviate this issue, we …

abstract arxiv cs.cv dynamic feature features global information memory nature prior reading reference representation template tracking type videos visual