Pseudo-Segmentation Guided Detection Refinement for Remote Sensing Small Object Tracking

Multi-object tracking (MOT) shines in the fields of remote sensing video processing and signal analysis. However, the dense and small objects of interest are prone to the failure of generic detection models and tracking strategies, thereby causing missing targets and ambiguous association. To address this issue, we propose the Pseudo-Segmentation guided detection refinement (PSTracker) for small objects in remote sensing tracking scenarios. Specifically, we introduce the simple yet effective pseudo-segmentation module to augment and improve localization information without complex pixel-level annotation. Subsequently, we present the additional long-term predictions which pose shape constraints on the matching conditions to accurately recall tracked targets with the large motion displacement. Additionally, we propose the corresponding multi-task joint learning to optimize the detection and pseudo-segmentation branches. Experimental results on two remote sensing benchmarks achieve the state-of-the-art performance in terms of the detection and association capabilities, which demonstrate the effectiveness and robustness of our tracker.