객체 검출 성능 향상을 위한 Extended paththrough layer 기반의 네트워크

Abstract

Object detection methods based on fully convolutional neural networks (FCNN), such as single shot multiple box (SSD), real-time object detection (YOLOv2), and region-based fully convolutional networks (R-FCN), achieves top performance in recent years. Among these methods, YOLOv2 has simple and fast network architecture, but it gives relatively lower detection performance for small objects. In this paper, we propose a object detection method based on YOLOv2 architecture to improve detection performance with reduced number of parameters. The previous YOLOv2 method increases the number of parameters by using paththrough layer which concatenates two feature maps of different levels. complexity by mixing high level feature and middle level feature. Our method improves both the detection performance and memory efficiency by removing the previous paththrough layer and adding extended paththrough layer.