Some questions about your loss functions.

[Wenretium]

Hello, I have two questions about your loss functions. In your paper, the Illumination-guided noise estimation loss is shown as: But in your codes, wn(x) represents illumination and wr(x) is also computed using illumination instead of the whole image I(x).

def reflectance_smooth_loss(image, illumination, reflectance): gray_tensor = 0.299image[0,0,:,:] + 0.587image[0,1,:,:] + 0.114image[0,2,:,:] gradient_gray_h, gradient_gray_w = gradient(gray_tensor.unsqueeze(0).unsqueeze(0)) gradient_reflect_h, gradient_reflect_w = gradient(reflectance) weight = 1/(illuminationgradient_gray_hgradient_gray_w+0.0001) weight = normalize01(weight) weight.detach() loss_h = weight * gradient_reflect_h loss_w = weight * gradient_reflect_w refrence_reflect = image/illumination refrence_reflect.detach() return loss_h.sum() + loss_w.sum() + conf.reffactorch.norm(refrence_reflect - reflectance, 1)

So I want to know why the codes are different from the paper. Did you update it? Because Illumination-guided sounds like actually the illumination but not I(x).

The second question is why you compute I-(RS+N) and R-I/S as loss components at the same time. In my opinion, it's contradictory. Thanks.

[hello-trouble]

Hello, I have two questions about your loss functions. In your paper, the Illumination-guided noise estimation loss is shown as: But in your codes, wn(x) represents illumination and wr(x) is also computed using illumination instead of the whole image I(x).

def reflectance_smooth_loss(image, illumination, reflectance): gray_tensor = 0.299_image[0,0,:,:] + 0.587_image[0,1,:,:] + 0.114_image[0,2,:,:] gradient_gray_h, gradient_gray_w = gradient(gray_tensor.unsqueeze(0).unsqueeze(0)) gradient_reflect_h, gradient_reflect_w = gradient(reflectance) weight = 1/(illumination_gradient_gray_h_gradient_gray_w+0.0001) weight = normalize01(weight) weight.detach() loss_h = weight * gradient_reflect_h loss_w = weight * gradient_reflect_w refrence_reflect = image/illumination refrence_reflect.detach() return loss_h.sum() + loss_w.sum() + conf.reffac_torch.norm(refrence_reflect - reflectance, 1)

So I want to know why the codes are different from the paper. Did you update it? Because Illumination-guided sounds like actually the illumination but not I(x).

The second question is why you compute I-(RS+N) and R-I/S as loss components at the same time. In my opinion, it's contradictory. Thanks.

I am also puzzled about this issue in this paper. In addition, the wr(x) is computed using the gray version instead of the reflectance.