Mix-groups Attention Network for Object Detection

Abstract

With the development of deep learning, object detection has made great progress in the past few years, and it mainly involves the combination of object classification and localization within a scene. Effective convolution features are critical to improve detection for objects of different scales, but learning powerful features remains a challenging task due to scale variations, occlusion. In view of this fact, we introduce a lightweight Mix-groups attention network to solve these problems. Firstly, a novel group attention module is proposed, which can learn more effective part-level information. Furthermore, to enrich multi-scale representations, it is combined with channel attention to generate Mix-groups attention module, assigning weights to image features in spatial and channel dimensions. The new approach can better highlight key objects, eliminate background clutter, and directly support multi-scale feature learning. A number of experimental results show that the proposed module has achieved excellent improvement.