While there is no paper or document descirbing YOLOv8 model architecture and training strategies so far, I started to find some other papers utilizing YOLOv8 for transfer learning, see if they have mentioned a little more about what kinds of new stuffs or technologies have been involved in YOLOv8 compared with the previous YOLOs.
For the neck, similar to YOLOv5, YOLOv8 also uses the methods of FPN and PAN.
For the head, simlar to YOLOv6 and YOLOX, decoupled head is used.
Similar to YOLOv6, YOLOv8 is also a anchor-free object detector that directly predicts the center of an object instead of the offset from a known anchor box which reduces the number of box predictions, and that speeds up the post processing process.
YOLOv8 uses Soft-NMS which is a variant of the NMS technique used in YOLOv5. Soft-NMS applies a soft threshold to the overlapping bounding boxes instead of discarding them outright.
The loss function in YOLOv8 is:
This loss function includes the CIoU (complete IoU) loss proposed by Zheng et al. [22] as the box loss, the standard binary cross entropy for multi-label classification as the classification loss (allowing each cell to predict more than 1 class), and the distribution focal loss proposed by Li et al. [10] as the 3rd term.
2. Results
Performance of YOLOv8 Variants
Press enter or click to view image in full size
mAP Against Model Parameters or Time
By plotting results from different YOLOv8 sizes, it is observed that YOLOv8 outperforms YOLOv5, YOLOv6 and YOLOv7.
