Bounding Box Width Regression Issue for Text Detection on x-axis in YOLO

[eslamahmed235]

Issue: Poor Box Width Regression for Text Detection on x-axis

Description

I am experiencing poor box width regression for text regions along the x-axis while using YOLO for a layout object detection task. The bounding boxes predicted for elongated text objects are significantly narrower than the ground truth.

Environment Details

• Image Size: 1025x1025
• Input Inference Size: imgsz = 640
• YOLO Version: YOLOV9
• Dataset: DoclayNet Dataset pdf pages screenshot saved as PNG with fixed size 1025.

Steps to Reproduce

1. Train YOLO on a dataset with text regions of varying aspect ratios.
2. Set the input inference size to imgsz = 640 while the original image size is 1025x1025.
3. Observe the bounding box predictions for text regions, particularly for elongated objects.

Observed Behavior

• The predicted bounding boxes for text regions are consistently narrower along the x-axis than the ground truth.

Expected Behavior

• Bounding boxes should tightly fit the text regions, accurately capturing their width along the x-axis.

Request for Assistance

Could you provide guidance or suggest specific configurations that might resolve this issue? Additionally, if this requires model or preprocessing updates, I am happy to contribute by testing or implementing the suggested changes.

Attachments