N

[ghost]

Search before asking

• [X] I have searched the YOLOv8 issues and discussions and found no similar questions.

Question

Hi this is very strange. I trained a detect YOLOv8 and was able to get great results

but when I try to test the model like use predict and source the image for all images both validation and training it gives no detections. Can someone please help. I am training on a virtual environment where i did a fresh install of ultralytics.

[ghost]

HI I would appreciate a pointer here please

[glenn-jocher]

@kkahol-percipio hello,

I'd be happy to help you out. Could you please provide more details about the issue you're facing with YOLOv8? Information about any error messages, the context in which they appear, and your current setup would be particularly helpful to understand and diagnose the problem. The more specifics you can provide, the better.

Looking forward to your response.

Best regards.

[ChaosSamKo]

I am having the same issue. Great metrics during training and validation but loading the last or best weights and using "predict" or "model(image_path)" just gives random/bad results.

[glenn-jocher]

@ChaosSamKo hello,

Thank you for reaching out and reporting this issue.

The problem you're experiencing may be due to several reasons. YOLOv8 is known for providing impressive metrics during training and validation due to its training process design. However, using the 'predict' or 'model(image_path)' method might not yield the same results for a couple of reasons.

Firstly, it's possible that the model is overfitting to the training data, which would result in great metrics during training but not as good when using it on unseen data. Overfitting occurs when a model is excessively complex and it starts to learn the noise in the data rather than the actual patterns. Methods like regularization and dropout can help make your model generalize better.

Secondly, another common problem might be related inconsistency between the pre-processing steps during training/validation and prediction. Make sure to use the same image preprocessing (resize, normalization, etc) in both cases. Also, ensure that you're providing the right input shape while using the prediction method.

If these don't solve your issue, it could be beneficial to give more details about how you're calling the 'predict' or the 'model(image_path)' function and the pre-processing steps you're following. This could help in identifying the cause and finding a solution more precisely.

I hope these suggestions help resolve the issue you're facing and look forward to hearing from you soon.

Best regards.

[glenn-jocher]

@kkahol-percipio i'm glad to hear that you've resolved the issue with the predictions. It's an important note that sometimes models can output lower confidence predictions that might be filtered out when using the default threshold. Using a lower confidence threshold, as you've done with conf=0, can help reveal these otherwise-hidden detections. This is particularly useful for diagnostic purposes.

It's worth considering that while setting conf=0 showed all potential predictions, in practice, you would typically want to find an optimal confidence threshold that balances precision and recall for your specific use case, as a conf=0 would also let through many false positives.

Furthermore, the ordered class predictions you are seeing is indeed indicative of top-K results, and it can provide insights into the model's behavior and confidence regarding different classes.

As a reminder, please remember to adjust the confidence threshold (conf value) as needed for the balance between true and false positives in a production setting or more rigorous evaluations.

For any questions or further assistance, please don't hesitate to reach out to the Ultralytics team or the community.

Best regards.

[xbkaishui]

@glenn-jocher Hi, I have a question about the conf parameter. how to choose best conf balanced precision and recall? can we conside best mAP point as a best conf score ?

[glenn-jocher]

@xbkaishui hello,

Choosing the best confidence (conf) threshold is indeed an important step in finding a balance between precision and recall. The confidence threshold determines the minimum probability that a detection must have to be considered by the model. Here are some considerations to help you find an optimal value:

1. Validation Metrics: The mAP (mean Average Precision) calculated during validation accounts for various confidence thresholds. Observing the precision-recall curve can give you insights into the performance at different thresholds.

2. Threshold Selection: There's no one-size-fits-all value for the conf threshold as it largely depends on your specific application's requirements. Applications that require high precision may set a higher threshold to reduce false positives, while those that require high recall, like medical imaging, may accept a lower threshold to minimize false negatives.

3. Use of mAP: While mAP provides a good overall indicator of performance, it does not directly tell you the best confidence threshold to use. However, examining the mAP at different confidence threshold values can guide you to an appropriate range.

4. Experimentation: One practical approach is to plot the precision-recall curve for your validation dataset and look for a point that offers an acceptable trade-off for your needs.

5. Domain-Specific Needs: Consider the costs of false positives versus false negatives in your application. This can help you to weigh precision against recall accordingly.

6. Automated Search: A systematic approach could involve automatically varying the confidence threshold in small increments and evaluating the impact on whatever combined metric is most important to you (such as the F1 score, which balances precision and recall).

Remember to validate such adjustments with an unbiased dataset to ensure generalization and avoid overfitting to specific confidence levels.

Best regards.

[xbkaishui]

ok thanks

[glenn-jocher]

@xbkaishui you're welcome! If you have any more questions or need further assistance in the future, feel free to ask. Good luck with your projects and happy detecting! 😊

Best regards.