Contextual-YOLOV3: Implement Better Small Object Detection Based Deep Learning

Small object detection is an open challenge due to its limited resolution and information. Existing object detection pipelines can't meet the requirement of accuracy for small objects. In this paper, we aim to address small object detection problem by introducing contextual information in detector. For this purpose, we propose an improved algorithm for fusing context in YOLOV3 called Contextual-YOLOV3. Specifically, in order to improve the accuracy for small objects. Contextual-YOLOV3 constructs a contextual relationship matrix combined with the classification probability of YOLOV3 to replace the original classification probability. Meanwhile we use context-based filtering algorithm for optimal window selection instead of non-maximum suppression algorithm (NMS). Extensive evaluations on the challenging Tsinghua-Tencent 100K and National Grid's intelligent inspection dataset benchmark well demonstrate the superiority of Contextual-YOLOV3 in detecting small objects.