Fix yolov3-tiny mask

[developer0hye]

Fix yolov3-tiny mask

original yolov3-tiny cfg

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Updated anchor masks for YOLOv3-tiny configuration for improved detection performance. 🎭

📊 Key Changes

• Modified the mask values of the YOLO layer in the yolov3-tiny.cfg file from 1,2,3 to 0,1,2.

🎯 Purpose & Impact

• 🎯 Purpose: To adjust which predefined anchor box sizes (width and height pairs) are used during detection at different scales. This can optimize the model to better detect objects of certain sizes.
• 💥 Impact: Users may experience changes in detection accuracy for their use case, as the model slightly alters its focus on different object sizes. This might result in improved performance for some datasets and use cases.

[glenn-jocher]

@developer0hye ah this is a very interesting story! One user on this repo was unable to reproduce the reported mAP with yolov3-tiny, and it turns out that the reason was that the darknet authors made a mistake when training yolov3-tiny, they actually used an incorrect anchor sequence in their cfg file (1,2,3 in the last layer instead of 0,1,2). So to reproduce their mAP (and get good inference results with yolov3-tiny), we need to maintain their original incorrect anchor sequence.
#188 (comment)

We have verified this by running the following code with the 1,2,3 sequence (high mAP), vs the 0,1,2 sequence cfg (low mAP):
python3 test.py --weights weights/yolov3-tiny.weights --cfg yolov3-tiny.cfg

[developer0hye]

Okay~! Thank you for your comment!

[glenn-jocher]

@developer0hye you're welcome! If you have any more questions or need further assistance, feel free to ask. We're here to help! Thank you for your support and happy coding!