Make crop work the same for pil and tensor #3770
Conversation
| padding = [max(-left, 0), max(-top, 0), max(right - w, 0), max(bottom - h, 0)] | ||
| return pad(img[..., top:top + height, left:left + width], padding) |
There was a problem hiding this comment.
Let's only call pad if needed like here: https://github.com/pytorch/vision/pull/3333/files
| right = left + width | ||
| bottom = top + height | ||
|
|
||
| if left < 0 or top < 0 or right > w or bottom < h: |
There was a problem hiding this comment.
Is this condition for bottom correct? I would have imagined that it should be similar than the one for right (which uses a > instead of a <).
There was a problem hiding this comment.
I missed it. Fixed this.
Thank you!
test/test_transforms_tensor.py
Outdated
| @@ -188,6 +188,10 @@ def test_crop(self): | |||
| 'crop', 'RandomCrop', fn_kwargs=fn_kwargs, meth_kwargs=meth_kwargs | |||
| ) | |||
|
|
|||
| # Test transforms.functional.crop including outside the image area | |||
| fn_kwargs = {"top": -2, "left": 8, "height": 4, "width": 5} | |||
There was a problem hiding this comment.
Can you add more tests exercising other use-cases, like when top is negative and height + top is larger than the image size?
There was a problem hiding this comment.
Added tests considering all directions!
|
Thanks for the changes! Lint is failing in https://app.circleci.com/pipelines/github/pytorch/vision/8117/workflows/29ea8926-e7f7-448b-b705-53905de1de37/jobs/580130 with the following message , but once you fix it the PR is good to merge. |
Summary: * Make crop work the same for pil and tensor * Only call pad if needed in functional_tensor.crop * Fix top-left functional_tensor.crop * Update document for functional.crop * Add other test cases of functional.crop * Fix bug * Fixing formattter * Fix stylings Reviewed By: cpuhrsch Differential Revision: D28538755 fbshipit-source-id: d38117a80ac6985e3f6a42863ee16903b6f5cea3 Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>
In the
transforms.functional.crop, if the cropping area is out of the image, in the case of pil, the box will be padded by zeros, whereas in the tensor, only the inside of the image will be returned.See this notebook.
In this PR
functional_tensor.croppad if needed.