Some Improved Strategies of YOLOv3 Algorithm

摘要

To enhance the representational ability of object detection network and make it learn features more comprehensively, thereby improving the accuracy of algorithm, some improved strategies are proposed. Based on widely used YOLOv3, the spatial pyramid pooling is embedded to strengthen the ability of feature extraction in the local area, and the integrated dilated convolution and dual attention are used to improve the quality of feature expression. Additionally, the regression loss is calculated by CIoU to improve the localization effect of bounding boxes. Experimental results on the MS COCO (test-dev 2017) dataset show that the mAP of improved model increases by 8.8%, while the detection speed of algorithm is maintained at the original level.