BATrack: End-to-End Tracking with Band Attention for Hyperspectral Video
Gaowei Guo, Zhaoxu Li, Qiang Ling, Wei An · 2024
The hyperspectral video contains frames with a large number of spectral bands, which provide fine reflectance information for hyperspectral object tracking (HOT). However, due to the domain gap between hyperspectral and RGB, it is not possible to directly use pre-trained models based on RGB. To address this issue, we propose an end-to-end transformer framework with band attention for HOT (BATrack). Specifically, a band attention block is incorporated to learn the interdependencies between bands and produce weights for each band. Select the three bands with the highest weights and composite the hyperspectral image into a three-channel image. These three-channel images are fed into a deep color tracking network to determine the target location. In addition, we have designed a template update strategy to select a more reliable template. The experimental outcomes on hyperspectral datasets have revealed the efficacy and the benefits of BA Track.