Regression Capsule Network for Object Detection

In recent years, the emergence of capsule networks has brou- ght new solutions to problems in various fields of computer vision. Capsule networks show the superiority of traditional convolution neural networks by using vectors and fusing the relationships between their different dimensions. Though prior work on object detection has shown their brilliant result, there is still a long way to improve in the detection results of multi-target and multi-classification tasks. Therefore, we consider that the combination of capsule networks and object detection can effectively improve the result of the prediction model. In this paper, 1. Our model utilizes the backbone structure of retinanet, using dynamic routing algorithm to ameliorate the structure of prediction subnet. 2. We test our model on MScoco 2017 datasets. mAP increases 0.8% and recall increases 0.4%.