incorrect display

[zdfy31220]

I tested your code with pytorch0.4.1 in windows, but the fusion result cannot display correctly. Any idea what caused this issue and how to fix it?

[pandayuanyu]

Hi Bro, have you fixed this problem? I got the same result.

[YZG-gaiciyang]

I got the same result.,too.

[DesBonbons]

Have you all solved this problem? Please tell me.

[sirizhou]

I have the save problem， Have you all solved this problem? Please tell me.

[maithal]

I have the save problem， Have you all solved this problem? Please tell me.

Was this issue resolved? Please share how you were able to solve it?

[rxqH]

我有保存问题，你们都解决了这个问题吗？请告诉我。

这个问题解决了吗？请分享您是如何解决的？
debug进去看了一下，是那个sunner得版本问题，下载指定版本就运行

[CharlesShan-hub]

Here is mine.

I changed a picture and plot the YCbCr: I find it seams Y channel is good:

[CharlesShan-hub]

I FIX IT！！！

because it used transform.Normalize

We need to use inv_trans!

inv_trans = transforms.Compose([
    transforms.Normalize(mean=[0., 0., 0.], std=[1/0.229, 1/0.224, 1/0.225]),  # Recover the std
    transforms.Normalize(mean=[-0.485, -0.456, -0.406], std=[1., 1., 1.]),  # Recover the mean
])

---

This is my fixed inference fuction

def inference(model,im1,im2,opts):
    # Load the Image
    trans = transforms.Compose([
        transforms.ToTensor(),
        transforms.Resize((opts.H, opts.W), antialias=True), # type: ignore
        transforms.Normalize(mean = [0.485, 0.456, 0.406], std = [0.229, 0.224, 0.225])
    ])

    [im1, im2] = [path_to_ycbcr(im) for im in [im1,im2]]
    assert(im1.size == im2.size)
    [im1, im2] = [torch.unsqueeze(trans(im), 0)for im in [im1,im2]] # type: ignore
    [im1, im2] = [im.to(opts.device) for im in [im1,im2]]

    # Fusion
    model.eval()
    with torch.no_grad():
        f_y = model.forward(im1[:,0:1,:,:], im2[:,0:1,:,:])  # Inference
        [f_cb, f_cr] = weightedFusion(im1[:, 1:2], im2[:, 1:2], im1[:, 2:3], im2[:, 2:3])
        [f_cb, f_cr] = weightedFusion(im1[:, 1:2], im2[:, 1:2], im1[:, 2:3], im2[:, 2:3])
        fused = torch.cat((f_y,f_cb,f_cr),dim=1)
    
    # Reconstruct the Fused RGB Image
    inv_trans = transforms.Compose([
        transforms.Normalize(mean=[0., 0., 0.], std=[1/0.229, 1/0.224, 1/0.225]),
        transforms.Normalize(mean=[-0.485, -0.456, -0.406], std=[1., 1., 1.]),
    ])

    fused = inv_trans(fused)
    fused = ycbcr_to_rgb(transforms.ToPILImage()(fused[0,:,:,:]))

    return transforms.ToTensor()(fused)

You can find code at my project: https://github.com/CharlesShan-hub/CVPlayground

---

Also I used some self-designed function:

def path_to_ycbcr(path: str) -> Image.Image:
    """
    Load an image from the given path and convert it to YCbCr format.
    """
    image = np.array(Image.open(path))
    if len(image.shape) == 2:
        image = color.gray2rgb(image)
    image = color.rgb2ycbcr(image)
    return Image.fromarray(image.astype(np.uint8), mode='YCbCr')


def ycbcr_to_rgb(image: Image.Image) -> Image.Image:
    """
    Load YCbCr format Image and convert to RGB format.
    """
    image_np = np.array(image)*1.0
    image_rgb = color.ycbcr2rgb(image_np)*255.0
    return Image.fromarray(image_rgb.astype(np.uint8), mode="RGB")