Paper address: YOLOv6 v3.0: A Full-Scale Reloading
Code address: https://github.com/meituan/YOLOv6
Yolov6 v3.0 is Meituan’s update to yolov6 in the Year of the Rabbit in 2023, which mainly focuses on network architecture and Novel improvements were made to the training regimen. The authors summarize these improvements as follows:
- The Bi-directional Concatenation (BiC) module is proposed in the neck network to provide more accurate positioning information.
Insert picture description here
- The anchor-aided training (AAT) strategy based on anchor boxes is used without affecting the reasoning efficiency. During the training phase, the auxiliary branch and the anchor-free branch learn from independent losses while the signals are propagated together. Therefore, additional embedded boot information from the auxiliary branch is integrated into the main anchorless header. It is worth mentioning that auxiliary branches are removed at inference time, which improves accuracy without reducing speed.
- The author deepens yolov6 to have another stage in the backbone and neck network, which enhances the ability to still achieve state-of-the-art performance at high resolutions.
- A new self-distillation strategy is proposed to improve the performance of the yolov6 small model. The self-distillation calculation is shown in the figure below:
