Comet Image Logging Fix

[DN6]

When doing some additional testing on the Comet Logger I caught a couple of issues.

1. The images being logged to Comet are not in the native scale, while the bounding boxes are. This leads to bounding boxes being unaligned across images
2. Since YOLO runs a final validation run after the training is finished, images logged in the final epoch were being logged again in the validation run.

This PR fixes both these issues.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced Comet.ml logging in YOLOv5 with threshold parameters and image logging improvements.

📊 Key Changes

• 📈 Import of PIL library.
• ✨ Added logging of validation IOU and confidence thresholds to Comet.ml.
• 🖼️ Store logged image names to avoid duplication.
• 🔄 Switched to logging the native-scaled images directly using PIL, instead of processing them.

🎯 Purpose & Impact

• ⚙️ The addition of PIL allows for better image handling during logging.
• 📊 By logging validation thresholds, users gain insights into the model validation process.
• 🚫 Avoiding duplicate image logs helps in maintaining a clean and efficient logging environment.
• 🌁 Logging native-resolution images gives a more accurate visual representation of model predictions.
• 🛠️ Overall, these changes enhance the Comet.ml integration, providing users with a more detailed and streamlined experiment tracking experience.

[DN6]

Hi @glenn-jocher @AyushExel. I was preparing some more examples for the Comet integration and caught a couple of bugs on the image logging side. This PR is to fix them.

Would appreciate you taking a look. Thank you!

[glenn-jocher]

@DN6 Fixes look good. PR is is merged. Thank you for your contributions to YOLOv5 🚀 and Vision AI ⭐