- Annotate the data with roboflow and then export the dataset in Yolo canonical format
- Model training content is in the Training folder
- Two versions of Yolov8 and Yolov11 models were trained, and the effect of the two models was tested separately. The effect of yolo11 is better.
- See the Models folder for the model files.
We can see that using the SAHI algorithm can indeed improve the accuracy of small object detection.However, since the sahi algorithm is to slice and detect the content and then merge it, it takes a lot of time, which will affect the efficiency of real-time detection. So I decided to use the sahi algorithm only in some frames.