Figure 6. Improved network TSR-YOLO.
3.4. Anchor Boxes Using K-Means++ Clustering
The original YOLOv4-tiny model’s anchor boxes were obtained by clustering the COCO dataset [38] and the Pascal VOC dataset [39]. By analyzing these datasets, we found that the targets in these datasets were more different in size and shape from those in the traffic sign dataset, and the background of the traffic sign dataset was more complex. As a result, the original anchor box size was unsuitable for Chinese traffic sign detection task and may harm the model’s training results. Figure 7 shows samples from the Pascal VOC dataset and the CCTSDB2021 dataset.
Do'stlaringiz bilan baham: |