Detail preservation and feature refinement for object detection

Abstract

In recent years, significant achievements have been made in object detection thanks to development of deep learning. However, there are still some open problems such as the poor performance in small object detection, especially when the computing resources are limited. In this paper, we propose a single-shot detail-preserving detector with a multi-flow sub-network and a multi-connection module, which is built upon the one-stage strategy to inherit the computational efficiency. Specifically, we first design a detail-preserving backbone network to preserve image details critical for small object detection. Then for feature refinement, we propose a multi-flow sub-network to optimize low-level features for small object detection and a multi-connection module to fuse multi-grained information to enhance feature representation without significant extra computational cost. Extensive experiments on PASCAL VOC and MS COCO demonstrate that our detector achieves state-of-the-art detection accuracies with high computational efficiency. The proposed method with only 300 × 300 input size achieves 82.6% mAP on PASCAL VOC 2007 and 32.9% mAP on MS COCO, both with one Nvidia Titan X GPU.