dvlab-research / snr-aware-low-light-enhance Goto Github PK

hello

The files in both sid and smid datasets are in'. npy' format. How do you convert'. npy' files to '.jpg' format using Pyraw

thank you

          Thank you, after you help , i solved this before , but there is a new problem that is I can't close to your results(24.61).When I train and test in LOL dataset, the best final result is 22.90(use l_pix+l_vgg as you described in your paper) ,I'm really interested in your work , and I'd really appreciate it if you could help me.

Originally posted by @id4su in #10 (comment)

Hello author, it is a great honor to see your work. I would like to ask you some questions about the use of the data set, that is, the SMID data set I downloaded contains 202 pictures of SMID_Long_png and 20809 pictures of SMID_LQ_png. How to divide the training set and the test set？and how about SID ?

Could you provide script to convert raw images from sid dataset to rgb image using rawpy’s default ISP.
Then, size of raw images in Sid dataset was 42562848, however from npy the input image are 960512 size.
Are you resizing the image

Thank you very much for your work. We would like to cite your paper and make a fair comparison with your method. Can you provide the evaluation code?

这里的nf没有定义，却直接append了。而nf的定义却在后面。

非常不错的工作~
我想请问一下，文章中报道的指标是用多少epoch的训练结果测试的。

当我进行运行test.py 脚本时，出现这样的错误，照片时随机在网上找的，请问这个问题该怎么解决？？ 谢谢，是我图片的问题吗、我用其他的照片可以测试成功。

How to deal with this problem? Looking forward to your reply,thank you!

你好，请问为什么训练的Loss会那么大呢？

hello,thanks for you greet job.i learn from it so much.
But i find some bad case when i use command "python test_LOLv1_v2_real.py -opt options/test/LOLv1.yml" to test with CPU.
here are those visual artifacts:
1.square artifacts in picture 'low00719.png' and 'low00764.png'

2. more blur comparing with GT

like the pixel in doll's eye

my main problem is why the artifacts rise.
the second artifact may be a good research point.

Hi, when I run this project, I found that
sou_img = util.tensor2img(visuals['LQ'][2])
save_img = np.concatenate([sou_img, rlt_img, ill_img, gt_img, rlt_img3, rlt_img2], axis=0)

should be
sou_img = util.tensor2img(visuals['LQ'])
save_img = np.concatenate([sou_img, rlt_img, ill_img, gt_img, rlt_img3, rlt_img2], axis=0)

Otherwise, their shapes will not match. The sou_img is [600, 400], not [600, 400, 3].
The error report is
File "<array_function internals>", line 6, in concatenate ValueError: all the input arrays must have same number of dimensions, but the array at index 0 has 2 dimension(s) and the array at index 1 has 3 dimension.

Hi author, thank you for your code, in 'dataset_LOLv2_real.py', the 'crop image' operation is not used and replace with resize (the image is resized with [608, 400]), so I want to know whether the resize is better than crop in actual experiment (But 'crop image' is used in LOLv1 dataset). I'm looking forward to your reply, thanks!

Is there a recommented method to test the model on unpaired images?

Different from original SDSD datasets with dynamic scenes, we utilize its static version (the scenes are the same of original SDSD) 这里的dynamic 和static version的区别是什么？
另外SDSD_indoor的百度网盘验证码似乎是不正确的

It is a great honor to see your work. I would like to ask why the decoder output is the residual instead of getting the reconstructed image directly.

你好，请问这个数据集为什么下载时提示没有权限？

Equation 6 is the paper is

I'm confused by the element-wise addition here. Since the SNR map is used to mask out tokens (i.e. patches) with low SNR, shouldn't the element-wise add be the element-wise product here? I.e.:

$$
\operatorname{Softmax}\left(\frac{\mathbf{Q}{i, b} \mathbf{K}^T{i, b}}{\sqrt{d_b}} \circ \left(1-\mathcal{S}^{\prime}\right) \sigma\right) \mathbf{V}_{i, b}
$$

Hi! thanks for the great work, could you provide the network name?