A Large-scale Detection Algorithm and Application Based on YOLOv4

Abstract

In this paper, we addressed the problems of occlusion and crowding in large-scale object detection. First, large-scale detection is more complex and diverse than traditional object detection. The number of targets to be detected is larger and often clustered together. This will produce occlusion and dense detection problems, which brings a serious challenge to object detection. Secondly, current dominant object detection is rarely trained and inferred on large-scale labeled dataset, so it is unable to evaluate the performance of these detection models on large dataset. To solve the above problems, we propose L-YOLO large-scale object detection algorithm. We modified the structure of feature pyramid network, then the receptive field was increased by using four-scale detection. Next, we propose a new loss function designed specifically for large-scale scenarios, which keeps the prediction box that is not the target as far away from the target as possible. It prevents the fusion of adjacent boundary boxes in the inference process and improves the detection performance in the case of occlusion effectively. At last, we use a new non-maximum suppression rule to prevent suppression of the correct detection box during infer. We annotated new dataset for large-scale detection, retrained and evaluated our model. Experiments on our dataset show the superiority of our model. Compared to the original YOLOv4, our improved model increases 1.8% mAP.