SPA-Tracker: Spectral Prompt Adaptation for Hyperspectral Object Tracking

Hyperspectral object tracking (HOT) is a crucial research area in computer vision and remote sensing. Unlike RGB-based trackers, HOT can more accurately capture the intrinsic material characteristics of objects. However, the limited number of training samples in hyperspectral datasets presents challenges for developing robust hyperspectral trackers. To tackle this issue, prompt learning is utilized to enhance spectral information in the input space while leveraging prior knowledge from foundational trackers. In this paper, we introduce a novel Spectral Prompt Adaptation Tracker (SPA-Tracker) that learns valuable prompts from hyperspectral images and adapts prompt information with hyperspectral, even when labeled training data is scarce. Compared to the baseline, SPA-Tracker includes two distinct modules: 1) A Spectral Information Prompt (SIP) module that extracts useful spectral information to enrich the input; and 2) An efficient Embedded Adapter (EA) module that facilitates learning interactions between RGB and hyperspectral features, thereby minimizing reliance on pre-trained RGB trackers. Extensive experiments demonstrate that our method achieves outstanding performance on HOT tasks.