Computer vision: how to create proper bounding boxes

Question

I'm creating a dataset to train a Yolov5 model to recognize tabular data. I want to clarify bounding box best practices. Would differences in bounding box tightness in the following two images affect the model's mAP?

Somewhat tight bounding boxes:

Tight bounding boxes:

Original image:

Since it takes more time to create the tighter boxes, I want to check whether it's worth the extra effort.

Answer 1

A precise and tight bounding box always helps in focus and learning more relevant features specific to object of interest. In this scenario, tight bounding box annotation would help in following ways:

• learning text edge based object bounding box and ignore the background features which will also help in reducing bounding box overlap with nearby text lines.
• Annotated bounding box will also make the model learn the text line spatial shape features(width, height)

Check this link to understand best practices for object detection annotation https://nanonets.github.io/tutorials-page/docs/annotate