YOLOCa: Center aware yolo for dense object detection

Abstract

Listen

Translate article

In this report, we propose a fully convolutional one-stage object detector (YOLOCa) with a novel decoupled head, an improved label assignment strategy, and a new loss computation method for dense object detection. By comprehensively considering the IoU, score and position connection between the prediction frame and the ground truth in the sample allocation, and by introducing some current mainstream loss function calculation methods such as Focal loss, DIOU nms, etc. Through the above techniques, we have effectively improved the AP. We conduct experiments on the global wheat detection dataset, and as shown in figure 1, our method achieves an AP of 49.9%, which is a 14.5% improvement over the baseline yolov3 (35.4%). This is still 0.8% AP higher than yolox which also uses the advanced label assignment strategy, and we still maintain the speed efficiency of the first-level object detector. We hope our work can provide some useful tips for developers and researchers, the source code is at https://github.com/gaozixiang13/YOLOCa.