Introduction:
YOLOv8 is the next major update of YOLOv5, which was open sourced by ultralytics on January 10, 2023. It currently supports image classification, object detection, and instance segmentation tasks. However, ultralytics did not directly name the open source library YOLOv8, but directly used the word ultralytics. The reason is that ultralytics positions this library as an algorithm framework rather than a specific algorithm. One of the main features is scalability.
Github address: https://github.com/ultralytics/ultralytics
network structure
Detailed explanation of core principles:
1. Backbone:
The idea of CSP is still used, but the C3 module in YOLOv5 is replaced by the C2f module with richer gradient flow. In order to reduce the weight, the number of blocks of the largest stage in the backbone network is also reduced, and the scaling factor N is different. The /S/M/L/X models no longer share a set of model parameters. The M/L/X large model also reduces the number of output channels of the last stage, further reducing the amount of parameters and calculations.
2. Neck:
There is no doubt that YOLOv8 still uses the idea of PAN, but by comparing the structure diagrams of YOLOv5 and YOLOv8, we can see that YOLOv8 deletes the convolution structure in the PAN-FPN upsampling stage in YOLOv5, and also C3 Module replaced with C2f module.
3. Head:
The current mainstream Decoupled-Head decoupling head structure is replaced, which separates the classification branch and the positioning branch, and alleviates the inherent conflict between classification and regression tasks.
At the same time, it was changed from Anchor-Based to Anchor-Free.
4. Loss:
TAL (Task Alignment Learning) dynamic matching is adopted, and DFL (Distribution Focal Loss) combined with CIoU Loss is introduced as the loss function of the regression branch, which makes the classification and regression tasks have a high consistency.
The matching strategy of TaskAlignedAssigner is simply summarized as follows: Select positive samples according to the scores weighted by classification and regression scores.
5. Training strategy:
The data enhancement part of the training introduces the operation of turning off the Mosiac enhancement in the last 10 epochs in YOLOX, which can effectively improve the accuracy.
Summarize
Looking back on the entire Year of the Tiger, it can be called the first year of YOLO involution. Meituan released YOLOv6, Alexey released YOLOv7, Baidu released PP-YOLOE+, Ali released DAMO-YOLO, and OpenMMLab released RTMDet. At the same time, these series of models are constantly being updated and iterated.
YOLOv8 mainly refers to the related designs of recent algorithms such as YOLOX, YOLOv6, YOLOv7 and PPYOLOE. There are not many innovations in itself, and it is biased towards engineering practice. The main promotion is the ultralytics framework itself.
