Abstract
Vehicle detection in satellite videos is challenging because of the small size of vehicles, low contrast, and complex backgrounds. In sparse detectors trained with center-based supervision, output nodes may be present within a vehicle’s bounding box while none exist at its center, leaving these useful nodes without appropriate positive supervision. To address this supervision mismatch, we propose FSSA-Det, a fixed-support supervision adaptation framework that improves vehicle detection by adjusting supervision at existing sparse nodes. The Support-Aligned Appearance Residual (SAAR) module samples local image features at the output coordinates to enhance sparse spatiotemporal features. When a vehicle center has no corresponding output node, Fixed-Support Supervision Adaptation (FSSA) removes conflicting negative heatmap supervision from selected in-box nodes and provides supplementary positive supervision to geometrically matched candidates. Positive–Negative Candidate Discrimination (PNCD) distinguishes vehicle candidates from high-response background candidates through response ranking and background suppression. On the seven VISO test videos, FSSA-Det achieves a macro-averaged F1 score of 81.63%, exceeding the baseline by 4.82 percentage points. Ablation studies further show that FSSA improves vehicle detection recall by adapting supervision at existing output locations.