A line-segment-based non-maximum suppression method for accurate object detection

Abstract

Computer vision models are currently making great strides in object detection with the rapid development of deep convolutional detectors. However, generating a large number of anchors is an indispensable step in the object detection models for locating targets, which inevitably leads to redundant detections and low computational efficiency. Detecting contours in an image is a fundamental cognitive ability in human vision system, which offers effective evidences for object detection. This paper proposes a novel and simple method by utilizing the distribution of line segments to facilitate the Non-Maximum Suppression (NMS) for the object detection models. Multiple differentiated metrics are designed for the overlap measure between bounding boxes. As a post-processing technique, the proposed segment-based NMS can be easily applied by various models. Furthermore, the proposed method is verified on multiple benchmarks and extensive experiments have been implemented to illustrate its effectiveness.