the-learning-and-vision-atelier-lava / passrnet Goto Github PK

我在demo_test.py 中测试了SSIM，（ SSIM = measure.compare_ssim(HR_left_np,SR_left_np,multichannel=True,data_range=1)）skimage用了0.13.0版本，但是跑出的middlebury数据集下的SSIM结果比论文中偏小。跑出结果middlebury: 0.822，论文结果middlebury: 0.871
请问有没有哪位正确测试出论文中middlebury数据集下的SSIM值？skimage用了那个版本？

Hi, Longguang, you have metion about the supplemental material many times in your paper. Where could I get it ?

Hello, how can I test without using the image on the right? How to set up the different test modes in Table 2?

您好！请问训练集是由Flickr1024中的800对训练图片和Middlebury中的60对训练图片构成的吗？

Hello,
In the project, the formula for calculating PSNR is as follows:
cal_psnr(HR_left[:,:,:,64:], SR_left[:,:,:,64:])

What does 64 represent?

Thank you.

Hello, how do you test with your own trained model? The model file you regenerate is in .tar format, for example: PassRnet_x4_epoch80.pth.tar. Is it necessary to change the model generated by myself to the model folder with the same name of the existing model file PASSRnet_x4.pth, and then test it?
Another problem is that you can only test a single test. When you put all the tests into the same dataset, can you test them at the same time?

您好，我正在学习您的这篇论文，能否提供一下论文测试集，谢谢！

您好！能提供X2和X4的预训练模型吗？

Hello，in the paper yousaid if the groundtruth disparities are available, we can generate the groundtruth attention maps accordingly（see the supplemental material for more details).
But the paper i downed didn't have the supplemental material,could you send me the material?
Thank you for your kindness!

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

Hi, thanks for your repo and it helps a lot.
I am trying to reproduce your repo recently and hit some bugs when testing my pre-trained model.

I would appreciate it if you could give me some advice.

When will you release the code?

我按照README中的操作步骤运行
python train.py --scale_factor 4 --device cuda:0 --batch_size 32 --n_epochs 80 --n_steps 30
系统给出如下错误：

Traceback (most recent call last):
File "train.py", line 107, in <module>
main(cfg)
File "train.py", line 103, in main
train(train_loader, cfg)
File "train.py", line 80, in train
psnr_epoch.append(cal_psnr(HR_left[:,:,:,64:].data.cpu(), SR_left[:,:,:,64:].data.cpu()))
File "/home/xujialang/PASSRnet/utils.py", line 84, in cal_psnr
return measure.compare_psnr(img1_np, img2_np)
File "/home/xujialang/anaconda3/envs/passrnet/lib/python3.6/site-packages/skimage/measure/simple_metrics.py", line 65, in compare_psnr
return peak_signal_noise_ratio(im_true, im_test, data_range=data_range)
File "/home/xujialang/anaconda3/envs/passrnet/lib/python3.6/site-packages/skimage/metrics/simple_metrics.py", line 149, in peak_signal_noise_ratio
"im_true has intensity values outside the range expected for "
ValueError: im_true has intensity values outside the range expected for its data type.  Please manually specify the data_range

我将'cuda:0' 切换为 'cuda:1' 或者其他GPU，错误就消失了，请问这是什么原因呢？

Hi，thanks for your great repo. I'm a little confused about why the left and right branches share weights during training. How can the same network extract different features from two different views? Looking forward to your reply. Thank you very much!