Addressing scale imbalance for small object detection with dense detector

Abstract

There are severe challenges on small object detection when using general object detector, especially scale imbalance on samples and features. Anchor-based detector performs poorly on small object detection because IoUs are too low to regress the objects. Key-point detector is hard to detect small objects for containing little semantic information. In contrast, a dense detector which directly learns an object via dense points is potential for small object detection. But they only catch fewer small objects because few of them can address scale imbalance brought by label assignment and feature extraction. We propose a dense detector named Libra EBox which includes Libra Ellipse Sampling (LES) and Residual Low-level Feature Enhancement (RiLFE). LES regulates positive regions of various objects by rescaling an ellipse box to catch samples of small object as much as possible. RiLFE is designed by several low-level feature maps to enhance small object’s feature representation. Experimental results show that our Libra EBox outperforms FoveaBox, FCOS and RetinaNet by 2.0% AP, 1.7% AP and 2.2% AP respectively for small objects on MS-COCO test-dev, and also outperforms most popular dense detectors on VisDrone-DET2018 and TinyPerson.