box candidates thresholds in hyp

[ml5ah]

@glenn-jocher
In response to issue #2521

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced augmentation filtering with new hyperparameters for better training stability.

📊 Key Changes

• Added three new hyperparameters to the hyp.finetune.yaml and hyp.scratch.yaml configuration files: ar_thr, area_thr, and wh_thr.
• Integrated these new hyperparameters into dataset augmentation functions within datasets.py.

🎯 Purpose & Impact

• 🎯 The addition of ar_thr (aspect ratio threshold), area_thr (area threshold), and wh_thr (width/height threshold) helps to filter out unrealistic bounding box transformations during data augmentation.
• 🏎️ These filters aim to improve the stability and effectiveness of model training by avoiding training on poor-quality or highly distorted images that could hamper the learning process.
• 🛠 Users can expect potentially more accurate object detection models when these filters are applied during the training of YOLOv5 on custom datasets.

[ml5ah]

Hey, @glenn-jocher checking to see if you got a chance to see and had any feedback on this?