CIoU protected divides

[glenn-jocher]

Protected divides in IOU function to resolve #8539

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Optimization of bounding box intersection-over-union calculations in YOLOv5.

📊 Key Changes

• 📐 Removed an epsilon (eps) addition in the width and height calculations of bounding boxes.
• 🔍 Adjusted the v term in the Complete Intersection over Union (CIoU) loss calculation to prevent division by zero during the aspect ratio term computation.

🎯 Purpose & Impact

• 🧹 The changes serve to clean up the intersection-over-union (IoU) computation, potentially leading to more accurate bounding box evaluations.
• 👨‍💻 For users and developers, these tweaks might result in slight performance enhancements during model training and evaluation, and more precise object detection metrics.

[UnglvKitDe]

@glenn-jocher Looks good! What about lines 232/233 and 250? There, it is also possible to divide with 0. Isn't it?:

yolov5/utils/metrics.py

Lines 232 to 233 in 7dafd1c

	inter = (torch.min(b1_x2, b2_x2) - torch.max(b1_x1, b2_x1)).clamp(0) * \
	(torch.min(b1_y2, b2_y2) - torch.max(b1_y1, b2_y1)).clamp(0)

yolov5/utils/metrics.py

Line 250 in 7dafd1c

return iou - (rho2 / c2 + v * alpha) # CIoU

[glenn-jocher]

@UnglvKitDe 233 is not a divide, the \ symbol means line continuation. The c2 on L250 is already protected by the + eps when it is defined in L244:

yolov5/utils/metrics.py

Line 244 in 7dafd1c

c2 = cw ** 2 + ch ** 2 + eps # convex diagonal squared

[UnglvKitDe]

@glenn-jocher You are right, srry didn't read correctly.

[UnglvKitDe]

@glenn-jocher Thanks for the quick fix!

[glenn-jocher]

@UnglvKitDe PR is merged. Thank you for your contributions to YOLOv5 🚀 and Vision AI ⭐