alterzero/RBPN-PyTorch

I find that you update the code,does this code fix "All the pixels are invalid in estimation Laplacian noise!!!"problem?

gittigxuy opened this issue · 9 comments

gittigxuy commented
I find that you update the code,does this code fix "All the pixels are invalid in estimation Laplacian noise!!!"problem?
aligoglos commented

I test latest clone and nothing change!
I have this problem and also my result resolution is same as input (size not increase)

  • 👍4
greatlog commented

I test latest clone and nothing change!
I have this problem and also my result resolution is same as input (size not increase)

I come up with the same problem. Have you solved it?

CVHW commented

I come up with the same problem.

Obliviously commented

I changed to Python 3.5.7 and rebuild pyflow with it. It seems to have solved the problem.

zhang-xi commented

In dataset.py，when the neighbor frame does not exist (line 35 and line 66), the same image as input will be appended into the array 'neigbor' (line 37 and line 68). Then pyflow will be applied to a pair of same images, which leads to the indicated problem. The simplest solution is to remove the first and last a few image file names in the list of training or evaluation files (such as Vid4/foliage.txt). The number of image files to be removed depends on the value of nFrames.

  • 👍1
zhang-xi commented

I test latest clone and nothing change!
I have this problem and also my result resolution is same as input (size not increase)

Maybe because the original image is down-sampled first by the factor of upscale_factor and fed into the model as input. Then the model will output the upscaled image which has the same size as the original image. Please check the code in dataset.py, on line 22 and 49.

aligoglos commented

I test latest clone and nothing change!
I have this problem and also my result resolution is same as input (size not increase)

Maybe because the original image is down-sampled first by the factor of upscale_factor and fed into the model as input. Then the model will output the upscaled image which has the same size as the original image. Please check the code in dataset.py, on line 22 and 49.

yes the issue of size related to these.

houqian2180320171 commented

In dataset.py，when the neighbor frame does not exist (line 35 and line 66), the same image as input will be appended into the array 'neigbor' (line 37 and line 68). Then pyflow will be applied to a pair of same images, which leads to the indicated problem. The simplest solution is to remove the first and last a few image file names in the list of training or evaluation files (such as Vid4/foliage.txt). The number of image files to be removed depends on the value of nFrames.

Hello, can you tell me more? Best to explain the drawing

chuan1093 commented

This is my solution:
In the file pyflow/src/OpticalFlow.cpp(about line 846), we can see why the error messages are printed: the content of the two input images are totally the same. And we can also see that OpticalFlow.cpp manages such a problem with a small number("para[k] = 0.001"), which is a proper solution. So I just comment the two lines of cout in order not to disgust me.
PS: remember to do this again:

cd pyflow/
python setup.py build_ext -i
cp pyflow*.so ..
  • 👍2