Hi guys, Thanks for the nice paper and the code. I tried to reproduce the button results in Fig. 6 from the paper without success. I am just trying to run an inference with the pretrained EDVR model but faced the following problems:

1. In fig. 6 you show x2 upscale but the model outputs the same image size as LR! I presume that this should act as a sharpening model, but is there another model for x2 upscale that you show in the paper?
2. After applying output = data_util.ycbcr2bgr(output), the result is bgr. I converted it to rgb in order to save as png with cv2.imwrite but the result is very poor and far from the rather sharp outcome that you present in the paper.

Will appreciate your answer.

Hi @IanYeung Given the YCbCr format of RealVSR Dataset, how could I visualize the Y channel and chrominance channel like you did in your paper? Could you please give me some suggestions? Thx a lot.

In the paper and program implementation, after decomposing the luminance channel, the charbonnier loss is used to calculate the loss, the selection mode is sum, and each decomposition loss is added. If the mode is chosen to be mean, the different layers of the pyramid cannot be summed due to inconsistent sizes. I chose mean and upsampled the pyramid losses of different scales, and added them after the scales were consistent. From this, the return tensor of the mean type is constructed. But the final reconstruction effect is that there are many regular small dots on the image. Do you have any comments?

I want to know what l_d_fake and l_d_real mean, l_g_gan is the sum of discriminator and generator? And what the 'd' and 'g' in l_d_fake and l_g_pix_s stand for?

Can you tell me how you convert RGB to YCbCr. The results I get with OPENCV conversion function are inconsistent with yours. Or can you show your code of RGB to YCbCr?

in test_RealVSR_wo_GT.py

dataset

read_folder = '/home/yangxi/datasets/RealVSR/test_frames_YCbCr'

the test_frames_YCbCr' in this path where did it come from？There is no file with this name in the downloaded dataset

How can I solve this problem when I run it?

(realvsr) lzw@computer3:~/code/RealVSR/RealVSR-main/codes$ python test_RealVSR_wi_GT.py 22-04-06 17:05:11.072 - INFO: Data: RealVSR - /home/lzw/code/data/ycbcr/LQ_YCbCr_test 22-04-06 17:05:11.073 - INFO: Padding mode: replicate 22-04-06 17:05:11.073 - INFO: Model path: /home/lzw/code/data/ycbcr/RealVSR-Models/001_EDVR_NoUp_woTSA_scratch_lr1e-4_150k_RealVSR_3frame_WiCutBlur_YCbCr_LapPyr+GW.pth 22-04-06 17:05:11.073 - INFO: Save images: True Traceback (most recent call last): File "test_RealVSR_wi_GT.py", line 220, in main() File "test_RealVSR_wi_GT.py", line 98, in main imgs = data_util.read_img_seq(subfolder, color=color) File "/home/lzw/code/RealVSR/RealVSR-main/codes/data/util.py", line 116, in read_img_seq img_l = [read_img(None, v) for v in img_path_l] File "/home/lzw/code/RealVSR/RealVSR-main/codes/data/util.py", line 116, in img_l = [read_img(None, v) for v in img_path_l] File "/home/lzw/code/RealVSR/RealVSR-main/codes/data/util.py", line 95, in read_img img = img.astype(np.float32) / 255. AttributeError: 'NoneType' object has no attribute 'astype'

您好，请教几个关于数据对齐的问题 1、请问图像配准是直接用Real-SR项目中代码做吗，有没有做了什么改进. 2、Real-SR中提到图像要先经过ps进行镜头矫正后再进行迭代配准，请问本文数据是否也同样做了矫正呢？ 3、请问你们配准成功率大概多少？ 谢谢了

Hello, I have some questions about data alignment

1. Is the image registration done directly with the code in the Real-SR project?
2. In Real-SR, the images were import into PS to correct the lens distortion before the image registration. Does the data in this paper also has lens distortion?
3. What is the registration success rate? Thanks