Preprocess image without F.interpolate

vision
1

I’m doing image prepressing during training as below:

im_transform = transforms.Compose(
  [
      transforms.ToTensor(),
      transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
  ]
)

im = Image.open('/content/image.jpg')

# unify image channel to 3 here

im = im_transform(im)

# add mini-batch dim here

im = F.interpolate(im, size=(im_rh, im_rw), mode='area')

Then I’m passing it to model. I checked how it looks like before passing to the model.

im_np = im.numpy()
im_np = im_np.reshape((im_rh, im_rw, 3))
im_np = (im_np * 255).astype(np.uint8)
check_im = Image.fromarray(im_np)
check_im

The input and processed image is looking like this,
Input Image
Processed Image

So my question is, how can I get the same result like this without using F.interpolate for inference. I’m reading image using opencv. And also normalize it to bound values between -1 to 1.

2

The interleaving in the processed image is most likely created by a view or reshape operation using a wrong memory layout (channels-last vs. channels-first).
I assume this operation might cause the trouble:

im_np = im_np.reshape((im_rh, im_rw, 3))

Could you check the shape of im_np before applying the reshape operation? If the channels are in dim0, either change the reshaping opertation or permute the dimensions first.

3

Exactly, the problem is not with F.interpolation or cv2.resize . Actually, im_np.reshape is not the correct choice here. I did it by this way and it works very well.

im = cv2.imread('/content/image.jpg')
im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)

# normalize values to scale it between -1 to 1
im = (im - 127.5) / 127.5   

# find scale factor - fx & fy

# resize image
im = cv2.resize(im, None, fx = x, fy = y, interpolation = cv2.INTER_AREA)

# prepare input shape
im = np.transpose(im)
im = np.swapaxes(im, 1, 2)
im = np.expand_dims(im, axis = 0).astype('float32')

Thank you for help.!