Accurate prediction box custom vision

Question

I would like to know how to make the red prediction boxes more accurate.

I am extremely accurate in providing the exact region with the according label but I don't see the accuracy of the red box getting better.
Examples:

region marking plus labeling:

Quick test result

Desired test result accuracy

Hope to get some input from the community

Answer 1

Hi, thanks for clarifying. The following document describes how to improve custom vision model. Based on the information you provided above, my recommendation would be to:

• Be aware of input image requirements
• Provide at least 30 images per tag
• Add more images and balance data; retrain
• Add images with varying background, lighting, object size, camera angle, and style; retrain
• Use new image(s) to test prediction
• Modify existing training data according to prediction results (you can use prediction images for further training)

Hope this helps!

Answer 2

Thanks for answering. I already have tried most of the points you mentioned but I also see some things I haven't tried. I appreciate your help!