Learning Progressive Spatial Enhancement and Scene Adaptive Alignment for Robust RGBE Tracking

RGBE tracking is a critical task that leverages visible images (RGB) from visual sensors and event stream (E) from dynamic vision sensors, realizing continuous locating of target. Existing methods often simply integrate the features of the RGB and E frames while neglecting spatial contextual information and modality-specific characteristics, leading to tracking failure in complex scenes. To this end, we extend the conventional RGB tracker MDNet to RGBE tracker incorporating with progressive spatial enhancement and scene adaptive alignment, named MDNet+. Specifically, a Progressive Spatial Enhancement (PSE) module is first introduced, in which spatial features are refined through two key strategies: Local Position Refinement (LPR) and Cross-space Information Interaction (CII). This design enables the generation of enriched feature maps by enhancing the effectiveness of spatial feature representation. Subsequently, the refined features are fed into a Scene Adaptive Alignment (SAA) module, where fine-grained interaction and fusion of modality-specific information are effectively performed. Extensive experiments on two public RGBE tracking benchmarks (VisEvent and COESOT) demonstrate that MDNet+ outperforms the existing state-of-the-art RGBE trackers, enhancing the tracking precision and robustness under challenging conditions. Notably, compared to the SOTA tracker FAEFTrack, MDNet+ achieves the improvements of 2.3% in Success Rate (SR) and 0.8% in Precision Rate (PR). The code is available at: https://github.com/Ambrumax/MDNetplus.