Restore remap boxes #812
Restore remap boxes #812
Conversation
# Conflicts: # tests/tensorflow/test_models_zoo_tf.py
Codecov Report
@@ Coverage Diff @@
## main #812 +/- ##
==========================================
+ Coverage 96.01% 96.10% +0.09%
==========================================
Files 131 131
Lines 4941 4955 +14
==========================================
+ Hits 4744 4762 +18
+ Misses 197 193 -4
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
There was a problem hiding this comment.
Hi @Rob192, thank you for your PR.
I added a few comments because I am not sure I fully understood the purpose of the remap_boxes function since we moved to a new format of boxes where boxes are described by points and not angles. In any case, can you attach in the issue #800 the image of the document where the boxes are misplaced to help us understanding the bug ? 🙏 Thank you!
| orig_height, orig_width = orig_shape | ||
| dest_height, dest_width = dest_shape | ||
| mboxes = loc_preds.copy() | ||
| mboxes[:, :, 0] = ((loc_preds[:, :, 0] * orig_width) + (dest_width - orig_width) / 2) / dest_width |
There was a problem hiding this comment.
I am not sure I understand well the computation here: If we have relative coordinates [[x1, y1], [x2, y2], [x3, y3], ...] we just need to multiply by dest_width & dest_height to get the absolute points coordinates in dest_shape referential, or am I missing something ?
There was a problem hiding this comment.
Hi @charlesmindee, no you need to take into account the padding that was introduced by rotate_image. This is the main point of this function remap_boxes that is here to accommodate these differences.
There was a problem hiding this comment.
I added pictures for reference in #800
There was a problem hiding this comment.
Ok, thanks I see the point
| orig_height, orig_width = orig_shape | ||
| dest_height, dest_width = dest_shape | ||
| mboxes = loc_preds.copy() | ||
| mboxes[:, :, 0] = ((loc_preds[:, :, 0] * orig_width) + (dest_width - orig_width) / 2) / dest_width |
There was a problem hiding this comment.
Ok, thanks I see the point
charlesmindee
added
framework: pytorch
This PR is to address #800 as part of #788.
It restores the function
remap_boxesfrom #488 and adds a new functionality test to verify the output of a pretrained predictor on a rotated document.