1

I trained Mask-RCNN model by Detectron 2 for only one class for instance segmentation of lines. If the line is located at an angle to the horizon close to 0 or 90 degrees and its bounding box does not intersect the boxes of other lines, then segmentation is performed correctly (left figure). On the central and right figures, the boxes of short lines at the corners do not intersect other boxes or intersect slightly and are also segmented correctly. But the masks for long lines are wrong. Are there any ways to fix this error? see figures

I trained 25000 iterations, loss_mask = 0.08, loss_box_reg = 0.19

Iurii Sobolev
  • 11
  • 2

1 Answers1

0

I think you should rotate an image untill the lines are at 0 or 90 degrees

John McClane
  • 1
  • 1
    thank you a lot. Unfortunately lines are not always parallel. Although the deviation from parallel is small, objects of the same class can still fall into the bounding box of one object. I think the error is due to this – Iurii Sobolev Feb 02 '23 at 05:54