Effective Feature Enhancement and Model Ensemble Strategies in Tiny Object Detection

We introduce a novel tiny-object detection network that achieves better accuracy than existing detectors on TinyPerson dataset. It is an end-to-end detection framework developed on PaddlePaddle. A suit of strategies are developed to improve the detectors performance including: 1) data augmentation based on scale-match that aligns the object scales between the existing large-scale dataset and TinyPerson; 2) comprehensive training methods to further improve detection performance by a large margin; 3) model refinement based on the enhanced PAFPN module to fully utilize semantic information; 4) a hierarchical coarse-to-fine ensemble strategy to improve detection performance based on a well-designed model pond.