Training models for multiple subclasses from a general class

[AdwaitMD]

Hi,
I have a specific problem. First, I want to train the yolov8 pose model to detect small objects (e.g. goats) from 5K images. then, I want to use a custom tracker and track the objects across the entire video. (I have already done this). In the next step, I want to extract crops of the different tracks/ objects, and train the yolov8-pose again to detect individual objects separately as different classes (e,g. goat 1, goat 2, goat 3) [All the objects are slightly different looking]. How can I do this? What is the best way to create training data out of crops?

[pderrenger]

Hi there,

Great to hear about your progress with YOLOv8! To achieve your goal of training the YOLOv8-pose model to detect individual objects as separate classes, you can follow these steps:

1. Extract Crops: Use the object cropping functionality to extract individual objects from your tracked video. You can refer to the Object Cropping Guide for detailed instructions and code examples.

2. Label the Crops: Once you have the cropped images, manually label each crop with its respective class (e.g., goat 1, goat 2, goat 3). Tools like LabelImg can be helpful for this task.

3. Prepare the Dataset: Organize your labeled crops into a dataset format compatible with YOLOv8. Ensure your dataset YAML file correctly references the paths to your images and annotations.

4. Train the Model: Use the YOLOv8 training script to train your model on the new dataset. You can start with a command like:

   yolo pose train data=your_dataset.yaml model=yolov8-pose.pt epochs=100 imgsz=640

By following these steps, you should be able to train your YOLOv8-pose model to recognize individual objects as distinct classes. If you need further assistance, feel free to ask! 😊

Best of luck with your project!

Kind