yhjo09 / sr-lut Goto Github PK

I would like to ask if we can change the fill position instead of rotation by rotating the image 3 times on the right and rotating it back to the original position after interpolation. For example, the previous filling on the right side was changed to filling on the upper, lower, and left sides respectively, and then the interception was corresponding to the previous one. I tried not to rotate, but my experiment found that the result seemed to be wrong. Is there any special significance of rotation.
millions of thanks.

In addition to changing the scale parameter during octuple upsampling training, does the network mechanism also need to be modified accordingly?

该LUT是单通道的？将输入图像分为三个通道，每个通道都查找LUT？我看到网络结构的输入也是单通道的

it's kernel size si smaller than A, why it's lut size is obviously larger than A?

Hello.
Thanks for your work.

Hi @yhjo09
I wanna to know that this LUT model(.npy) is run on CPU or GPU?
if it is run on CPU, it is possiable to run on GPU for fast inference?

   Thank you very much! ^_^ 

Hi,
The method to increase the receptive filed in this paper, actually, is rotating and padding. I'm trying to use the padding alone to increase the receptive filed, specifcally, padding the original image at (top or down) and (left or right) four times, and ensemble the results. (for example in the picture below, yellow area results will be ensembled). But the performance even obviously worse than the method that do not ensemble.
It's strange, have you tried this method?

Thank you for your great work at first.

Now we are planing to run SR-LUT on a PC(i7 CPU), we are considering to convert Python project (inference part) to VS2017.

I think there will be no big problem for python converting to C++, but we concern the speed issue.

I studied issue #1，and know that we should use parallel execution to improve the speed. for Android, it is Stream API. But for VS2017 on PC. would you please give me some suggestion on how to do parallel execution ? Is multipe thread a good solution?

Thank you in advance.

Hi,

Thank you for sharing your code with the community!
I wanted to ask you if it's possible for you to share the implementation of the mobile application? Also which model is used in this later?

Many thanks in advance!
With best,

Hi, i tested the psnr value on Set14 using the lut provided in this repo Model_S_x4_4bit_int8.npy, and get the value 26.41, which is different from paper that is 27.01. I used python package PIL to generate low resoultion image of Set14 (bicubic mode).

So i'm curious about why there is such a big difference.

Thanks for your sharing, hope for your reply

No training code was found for model_F and model_V. Could you please provide the relevant code？

Is there anyone could help me？

hi，@yhjo9 According to the script provided by the warehouse, psnr has been oscillating around 27.5 (about 400,000 times) during the sr-lut training process. Is this the final result that the model can achieve? thanks

In Train_Model_S.py Line 182-184, each output is clip to [-1, 1], then the batch_S is projected to [-2, 2], but the batch_H belongs to [0,1].

Is this right to compute the L1 loss on [-2, 2] and [0, 1] ?

这个是1_train_model_s阶段，他大概持续了几个小时也没有反应但也没报错请问它在正常工作么？期待您的回复！

Hi @yhjo09
I think your SR-LUT method is very good.Can you provide the source code of SR-LUT.apk?I am very eager to learn about it.Thank you very much.

We can't achieve these speeds on the test set, which can take hundreds of milliseconds. When testing 512×256 images, the running time reaches tens of seconds. What Settings are required to achieve such a fast speed? My CPU is 12 vCPU Intel(R) Xeon(R) Platinum 8255C CPU @ 2.50GHz