Graph Convolution and Self Attention Based Non-maximum Suppression

Abstract

Non-maximum suppression is an integral and last part of object detection. Traditional NMS algorithm sorts the detection boxes according to their class scores. The detection boxes with maximum score are always selected while all other boxes with a sufficient overlap with the preserved boxes are discarded. This strategy is simple and effective. However, there still need some improvements in this process because the algorithm makes a ‘hard’ decision (accept or reject) for each box. In this paper, we formulate the non-maximum suppression as a rescoring process and construct a network called NmsNet which utilizes graph convolution and self attention mechanism to predict each box as an object or redundant one. We evaluate our method on the VOC2007 dataset. The experimental results show that our method achieves a higher MAP compared with the traditional greedy NMS and the Soft NMS.