yingqianwang / ipassr

[CVPRW 2021] Symmetric Parallax Attention for Stereo Image Super-Resolution

🌱News🌱

• 2022-03-02- Our paper "Occlusion-Aware Cost Constructor for Light Field Depth Estimation" is accepted by CVPR 2022. Code is available at OACC-Net.
• 2022-02-16- Our paper "Disentangling Light Fields for Super-Resolution and Disparity Estimation" is accepted by IEEE TPAMI.
• 2021-12-25- The source codes of our DistgSSR, DistgASR, and DistgDisp have been released.
• 2021-12-10- An online review of light field image super-resolution is available at LF-Image-SR, and an open source toolbox is available at ZhengyuLiang24/BasicLFSR.
  
  

ipassr's Issues

Last Convolution

Why have you put a convolution after the pixel-shuffle layer, although you didn't mention it in your paper?

About test on Flickr1024 dataset for 2x SR

Hi, thanks for your sharing! I want to test iPASSR_2x.pth on the Flickr1024 dateset for 2x SR using the test.py provided in iPASSR and demo_test.py provided by PASSRnet on a PC with NVIDIA GeForce RTX 3090. But it is always out of memory, is there any way to solve it? Looking forward to your reply. Thank you very much!

torch.cuda.OutOfMemoryError: CUDA out of memory.

Hi I ran into this problem:
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 10.73 GiB (GPU 0; 8.00 GiB total capacity; 17.20 GiB already allocated; 0 bytes free; 19.15 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
How can I solve this ?

dataset

First of all, I would like to thank the author for his work. I am new to this field. I would like to ask. After downloading the Flickr1024 and Middlebury data sets, I followed your request and executed the GenerateTrainingPatches.m code in IPASSR to generate image pairs, but found that the generated It is a mixed data set of Flickr1024 and Middlebury data sets. Then when training, how to distinguish which data set is being trained?

Comparison to Bicubic in another dataset

Thank you for your great repo.

I am trying to test your method on another dataset for stereo video SR and compare it with you.

I did not train your model on my dataset, and just used your pre-trained model to create super resolved frames.
The problem is that PSNR of your method is lower than the Bicubic base method. When Bicubic gives about 31, your method gives 28.5 of PSNR.

Why do you think this happens?
Thank you so much

Training Dataset

Thank you for the amazing work! I am unable to access the training data through baidu. Is the training data available on google drive? Thanks

what does is_training=1 and is_training=0 mean respectively

Some questions about the evaluation.

Excellent Work!! I recently use the code to conduct the stereo SR experiments, but I run the evaluation.m to test the performance and the code reports an error, as "conv2 cannot handle n-dimensional arrays, such as RGB images". Looking forward for your reply!

How do you train the model?

Hi, thanks for sharing.
I read your Stereo SR and LF SR code, but there are no validation code, so I want to know how do you find the best model(e.g. in iPASSRnet, the 80 epochs). Or, you just manually use the test code every epoch to do validation? BTW, when do you share the LFDFNet code? Is it possible to use the PAM module for LF SR ?
Again, thks a lot.

Waiting for your kindly reply,
Regards,

About crop 64 pixels in evaluation

Hi, thanks for your sharing! I wonder why the left view image needs to be cropped left 64 pixels for evaluation. It confuses me a lot, and I hope you can help me. Thank you for your kindness!