Tiny object detection based on dynamic scale-awareness label assignment and contextual enhancement

Abstract

The prosperity of recent object detection can not camouflage the deficiencies of tiny object detection. The generic object detectors suffer a dramatic performance degradation on tiny object detection. For this purpose, we present a tiny object detection approach based on Dynamic scale-awareness label assignment and Contextual enhancement (DCNet), which improves the tiny object detection performance from label assignment and feature enhancement perspectives. Considering the IoU-based label assignment seriously harms the positive samples for tiny objects, we design a Dynamic Scale-Awareness (DSA) label assignment to replace it in the region proposal network. The DSA label assignment adaptively rescales preset anchors and introduces the regression information to better assign the preset anchors for tiny objects. Furthermore, the tiny objects often exhibit weak feature responses due to their poor-quality appearance. Therefore, we propose a contextual enhancement module that aggregates contextual information at different scales to enhance tiny objects’ feature responses. Comprehensive experimental analyses on multiple datasets confirm the effectiveness and good generality of our proposed DCNet in tiny object detection.