A keypoint-based object detection method with attention mechanism and feature fusion

Abstract

Recently, there is a new object detection framework that does not require anchor boxes, which refers to the realization of object detection tasks by detecting key points. CenterNet identifies an object with single keypoint, namely the center point of its bounding box. It finds other attributes at the same time through key point estimation, such as the size and the orientation of the object. In this work, a global attention module is introduced to the backbone called Hourglass to enhance feature extraction with the global context information. A multilevel fusion method is also added to the Hourglass to integrate the feature maps of different levels, and further improve the detection capability. Combining the two methods, the new network achieves 46.1% AP with multi-scale testing on MS COCO.