SODBoost: Density-Guided Zoom-In Mosaic for Small Object Detection

Small object detection (SOD) underpins critical applications such as aerial surveillance, autonomous driving and robotics, yet remains inherently challenging due to weak feature representation and unstable optimization. To address these issues, we propose SODBoost, a plug-and-play framework designed to Boost Small Object Detection with two training-only modules. Density-Guided Zoom-In Mosaic targets the problem of weak supervision for small objects by adaptively focusing data augmentation on high-density small object regions, thereby enriching effective supervision without compromising normal-scale performance. Scale-Adaptive Similarity mitigates the instability of IoU-based regression by introducing a scale-aware interpolation between location similarity and GIoU through a gating mechanism, producing smooth and consistent gradients across scales. Integrated into representative real-time DETR-based detectors, extensive experiments on VisDrone benchmark demonstrate that SODBoost consistently enhances small object detection performance while incurring zero inference overhead.