Enhancing Small Object Detection via Uncertainty-Aware Modeling

Although object detection has achieved significant progress in recent years, current detectors have achieved robust performance primarily on general-sized objects. However, detecting small objects remains a major challenge due to their limited pixel information, poor features, and high localization ambiguity. These inherent difficulties lead to a lack of effective positive samples during training and limited feature representation for small targets. Prior works have attempted various improvements, such as refined label assignment strategies and multi-scale feature enhancement, but performance bottlenecks still persist. In this work, we propose an uncertainty-aware detection framework specifically designed to handle small object detection. Our approach first generates an uncertainty-strengthened response map that is highly sensitive to potential object regions. This map guides the assignment of more positive anchors to suspected object regions, increasing the density of valid training samples. Subsequently, we integrate this uncertainty response map with the backbone features to guide the detector in an uncertaintyaware manner. Experimental results on the AI-TOD dataset demonstrate that our method achieves an <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$A P_{\text{vt}}$</tex> improvement of 2.2%, outperforming the baseline and validating the effectiveness of our strategy.