Object Detection with rotated bounding box

vision
1

Hello everyone, I have a custom dataset for text detection. Most of the text is tilted so does the bounding box around it. I have checked that many algos like YOLO, SSD work good with rectangular bounding box but haven’t seen any example where these algos are used to detect rotated bounding box.

Since most of the text in the given dataset has a rotated bounding box, should I adjust the coordinates of those bounding box to make them rectangular or should I regress all 4 coordinates of bounding box instead of width, height and center coordinate of the bounding box as used in YOLO.
Kindly help me out here.
Thanks

2

there are mainly 2 ways to do this:

  1. Regression bbox and an angle which decribe the rotation of the box;
  2. Directly regression the 4 points as a polygon of the shape. these method have Centernet Cornernet etc can do so.
3

You can use an object detection algo which also predicts object orientation. Check out Localizer, this is a python library for object orientation detection.