F.crop behavior diverges between tensor and PIL images

[pmeier]

import torch
from torchvision.prototype.transforms import functional as F

image_tensor = torch.rand(3, 7, 33)
image_pil = F.to_image_pil(image_tensor)

print(
    F.get_spatial_size(F.crop(image_tensor, top=9, left=9, height=20, width=12)),
    F.get_spatial_size(F.crop(image_pil, top=9, left=9, height=20, width=12)),
)

[22, 12] [20, 12]

cc @vfdev-5 @datumbox @bjuncek

[pmeier]

Actually, this is not a prototype issue:

import torch
from torchvision.transforms import functional as F

image_tensor = torch.rand(3, 7, 33)
image_pil = F.to_pil_image(image_tensor)

print(
    F.get_image_size(F.crop(image_tensor, top=9, left=9, height=20, width=12)),
    F.get_image_size(F.crop(image_pil, top=9, left=9, height=20, width=12)),
)

[12, 22] [12, 20]

[vfdev-5]

Seems like a bug in padding_ltrb, it may have to be something like :

        padding_ltrb = [
-            max(-left, 0), 
-            max(-top, 0), 
+            max(-left + min(right, 0), 0),
+            max(-top + min(bottom, 0), 0), 
-            max(right - w, 0), 
+            max(right - max(w, left), 0), 
-            max(bottom - h, 0), 
+            max(bottom - max(h, top), 0),
        ]        

This will also fix the issue with too negative top-left corner and small crop size:

import torch
from torchvision.transforms import functional as F

image_tensor = torch.arange(5 * 5, dtype=torch.float).reshape(1, 5, 5)
image_pil = F.to_pil_image(image_tensor)

o1 = F.crop(image_tensor, top=-10, left=-10, height=3, width=3)
o2 = F.crop(image_pil, top=-10, left=-10, height=3, width=3)
o1.shape, o2.size
> (torch.Size([1, 10, 10]), (3, 3))