cig-skoltech / deep_demosaick Goto Github PK

Excuse me, I want to know how to train the denoiser before running the main.py to train the demosaicking model?

Thank you

Hi, this work is very interesting. When I testing the performance on my own captured image, I'm confused about the input shape of MMNet.

p = Demosaic(image_patch.float(), M_patch.float())
p.cuda_()
xcur = mmnet.forward_all_iter(p, max_iter=args.max_iter, init=args.init, noise_estimation=args.noise_estimation)

(1) In my testing, the size of image patch is (H, W) with rggb bayer pattern. The largest number of image patch should be 255.
(2) However, when I see the code in data_loader.py, the size of input seems to have size of H, W, 3.

image_mosaic = np.zeros(image_gt.shape).astype(np.int32)
image_mosaic[:, :, 0] = mask[..., 0] * image_input
image_mosaic[:, :, 1] = mask[..., 1] * image_input
image_mosaic[:, :, 2] = mask[..., 2] * image_input
#print(image_mosaic.dtype)
 image_input = np.sum(image_mosaic, axis=2, dtype='uint16')
# perform bilinear interpolation for bayer_rggb images
if self.apply_bilinear:
image_mosaic = self.preprocess(self.pattern, image_input)

image_gt = img_as_ubyte(image_gt)
image_input = image_mosaic.astype(np.float32)/65535*255

What's the right understanding of this question. Looking forward for your reply. Thank you.

Excuse me, your codes don't have the 'data' folder, so where and how can I download the dataset?

When running the test code written by myself, I can't obtain the results in your paper. Can you also release your test code? It would help a lot to check my mistake. Thank you

Excuse me, I want to know how to train the net-final.mat before running the main.py to train the demosaicking model? As if I don't load the resDNetPRelu_color_prox-stages:5-conv:5x5x3@64-res:3x3x64@64-std:[1,2,3,4,5,6,7,8,9,10,11,12,13,14,15]-solver:adam-jointTrain/net-final.mat before training the MSR dataset, the performance will drop dramatically. Could you please tell me how to train the net-final.mat?

Thank you