Asymmetrical Reverse Connection and Smooth-NMS for Object Detection

In this paper, we propose a new network structure, a more efficient object detection framework. Inspired by the original RON, we also joint the region-based and region-free methodologies of object detection. There is a lifting space in the accuracy of the original RON, so we design the following two structures: (a) design a new reverse connection structure, which can obtain much more information in small object detection; (b) design a new inception structure based on asymmetric convolution to improve the efficiency of object detectors. The conventional of non maximum suppression is replaced by more efficient Smooth-NMS in the object detection phase. With the use of low resolution 320 * 320 input size, the new network structure achieved 75.6% mAP (our method is 1.2% higher than the original RON) and 71.8% mAP on the standard PASCAL VOC 2007 and 2012 datasets respectively. The experimental results show that our method can generate higher detection accuracy.