We have a size mismatch when generating document assets that undergo 'document cleanup' step. This happens due to ensuring that the document is a modulo 32 .
Image of size 2560 x 2384 will become 2592 x 2400.
def preprocess(self, img: np.ndarray) -> np.ndarray:
# check if PIL image
if not isinstance(img, Image.Image):
img = Image.fromarray(img)
# make sure image is divisible by 32
ow, oh = img.size
base = 32
if ow % base != 0 or oh % base != 0:
h = oh // base * base + base
w = ow // base * base + base
img = img.resize((w, h), Image.LANCZOS)
# convert to numpy array
return np.array(img)
Not only that but we are performing an resize instead of overlay. So we are effectively modifying the original image.
Expectation is that after running segment_frame the input and output images will be of this same size.
We have a size mismatch when generating document assets that undergo 'document cleanup' step. This happens due to ensuring that the document is a modulo 32 .
Image of size 2560 x 2384 will become 2592 x 2400.
Not only that but we are performing an
resizeinstead ofoverlay. So we are effectively modifying the original image.Expectation is that after running
segment_framethe input and output images will be of this same size.