Multiple instances predicted over the same ground truth instance YOLOv8

[genaroNF]

Hello everyone and thanks in advance for your help!

We are using the YOLOv8 model and we are trying to segment some ant roads, the problem we are facing is that sometime some roads are detected multiple times. We searched for a solution and we understood that we need to modify the IoU Threshold when training by incrementing it since it will tell the model to focus on getting a higer surface of intersection over the ground turth.

Is there a way to modify the IoU threshold when training the YOLOv8 Model? maybe we missunderstood and it should only be modified when predicting.

Thank you!

[UltralyticsAssistant]

👋 Hello @genaroNF, thank you for reaching out and for using Ultralytics YOLOv8 🚀! Let's dive into your query about adjusting the IoU threshold.

It sounds like you'd like to modify the IoU threshold to handle instances where multiple predictions overlap with the same ground truth. You have two opportunities to adjust IoU-related parameters:

1. Training Phase:

   • During training, adjusting the IoU threshold can impact how the model handles positive and negative samples. You can modify this in the training configuration file by setting the iou_thres parameter.
   • Find your training configuration (usually a YAML file) and look for the section related to hyperparameters. You can modify the IoU threshold as follows:

   # hyperparameters section
   iou_thres: 0.45  # Default is 0.45, increase to more strictly evaluate overlaps

2. Inference Phase:

   • When predicting, you can also adjust the IoU threshold to affect non-max suppression (NMS). This will help filter out overlapping predictions. You can do this directly in the code or by using command-line arguments:

   model = YOLOv8.load('yolov8.pt')  # Load a pre-trained model
   results = model.predict('image.jpg', iou_thres=0.5)  # Adjust iou_thres as needed

   Or via the CLI:

   yolo task=detect mode=predict model=yolov8.pt source=image.jpg iou_thres=0.5

For best practices when submitting discussions or issues, please provide:

• Detailed description of the issue.
• Steps to reproduce, if applicable.
• Configuration details and any custom modifications.

👉 Check our Docs for more guidance on training and prediction. If you still face challenges, feel free to ask more questions!

This is an automated response, but an Ultralytics engineer will assist you soon. For real-time support, you can also join our Discord community 🎧.

Happy modeling! 🎉

[glenn-jocher]

@genaroNF IoU threshold is an inference argument, not a training argument. You can modify this when predicting with results = model.predict("path/to/image.jpg", iou=0.45)

[genaroNF]

Ok, but then what does iou_thres do if you set it in the config file? does it not apply to YOLOv8 anymore or something like that?

Thanks for your time!

[glenn-jocher]

@genaroNF the iou_thres in the config file is typically used for inference to filter overlapping detections. It doesn't affect training in YOLOv8.

[genaroNF]

Got it, thank you so much!

[glenn-jocher]

You're welcome! If you have any more questions, feel free to ask.