vandermode/errnet

Single Image Reflection Removal Exploiting Misaligned Training Data and Network Enhancements (CVPR 2019)

errnet's Issues

Regarding synthetic training dataset (7,643 images) for reflection removal

Hi,

How do I save the synthetic training dataset for reflection removal generated using PASCAL VOC2012 images locally? Is it being generated on the fly? Could you please provide code pointers / provide the synthetic training dataset? (mixed, background and reflection layer images).

Thanks
Rohan

Can it be used to remove watermarks?

core dumped reported when test mobile camera image

Dear team,

i capture Nexus 6P phone with less resolution image and same image use to evaluate model, but encounter Abort(core dumped) with following error on cpu, but model able to run smoothly on other data set. Any precondition or image process need to follow for own test image?

CUDA outofmemory error。

HI，thanks to your great work.
i want to test one image, and save the result. but meet the error:

why test one image need so big memory? or somthing wrong else?

Cuda runnign OOM

Dear Team,

Am facing "Cuda running OOM" problem while runing test few sample, am not able to solve this problem using "--gpu_ids 0,1" , please help me to solve this problem.

reset

Hi, I've noticed that your code runs reset () after each epoch, but this function is only for CEIL data, and there is no reset for real20. I'd like to ask a related question about this.
If I use all my data, just like real20, it has been synthesized in a fixed order, whether it is necessary to reset(), whether it is necessary or optional.
@Vandermode

Hi, does gradient operator loss improve results of reflection removal

Hi，I think you have not use gradient operator loss in your code, but in your paper it does

ImportError: cannot import name 'imresize'

I was trying to run ERRNet on my machine but was getting this error again and again.
I'm running
python test_errnet.py --name errnet -r --icnn_path checkpoints/errnet/errnet_060_00463920.pt --hyper
According to the official document, 'imresize' is removed in SciPy 1.3.0.

About the suppl. material

Could you release your supp? I read your paper and you refer the suppl. material, but I searched the CVPR 2019 Open Access Repository and the google website I cannot find the suppl, look forward to your reply.

Pretrained model

Hi, I used the pretrained model you provided to test your dataset. But it's not as good as your paper shows. Could you give me some advice on this problem? @Vandermode

do i need to make folder "/media/kaixuan/DATA/Papers/Code/Data/Reflection/VOCdevkit/VOC2012/PNGImages"

/ERRNet/data/image_folder.py", line 33, in make_dataset
assert os.path.isdir(dir), '%s is not a valid directory' % dir
AssertionError: /media/kaixuan/DATA/Papers/Code/Data/Reflection/VOCdevkit/VOC2012/PNGImages is not a valid directory

KeyError: 'real' when batch size > 1

Hi, it seems that this happens when the data in a training batch isn't from the same dataset. (caused by the extra 'real' tag in train_dataset_real)
Did you train with only batch size=1? In my case it takes 3GB and I think it could be faster with larger batch size.

sliding-window strategy

hi, thanks for your great work, when i test a large image, i got oom error.
And i have seen the old issues, do you provide sliding-window strategy in your code?

Pretrained model from Google Drive is not found

404 The requested URL was not found on this server.

Your pretrained weights have a problem

A few minute ago i did test on CEIL Synthetic dataset with your pre-trained weights.
But, When i check the results, As well as other quantitative results are low, all of inferences are black.

unaligned datas

Hi, I just trained the errnet with the aligned datas follow the second part (Playing with aligned data) on GitHub. And The effect is good on aligning datasets. And could you give me some advice on how to utilize misaligned training data?

Training problems

Hi, when I reproduced your results（only with aligned data）, I found that the loss for the first ten times fell very fast, and the subsequent decrease was slow. For example, VGG dropped from 4 to 1.81 (1.31 after 60 iterations) and Ipixel dropped from 0.021 to 0.0094 (0.09 after 60 iterations) after only 15 repetitions.
I also tested testdata_reflection_synthetic_table2 with the 15-iterations model and got relatively good visual effects. Is this normal? Or something i did wrong.

Taining image crop method

Dear Team,

To crop the center region with size 224 x 224 any specific technique is used to crop and save is acheived?

Does patch-gan helps to improve quality of reflection removal

Hi, I have read your code, and you have implemented patch-gan discrimator, but not used. I just doubt that if it helps to improve the quality of reflection removal

Result I got

input

output

I tried to test one single image using the way the below question told.
The result tend not to be well as the result in the paper. So is there anything i've done wrong during testing.

Pascal VOC dataset

Your work is very great. But we have some questions about Pascal VOC dataset. When we train, we don't find the ground truth of Pascal VOC dataset. Do we miss something, or how do you generate the GT of Pascal VOC?

Original results

Can you share the original result images? Thanks.

About the resize problem of real20

Hi，I've resized the short edge of images in real20 dataset to 512(used torchvision.transforms.Resize).. I tested your released model but it got psnr of 21.89, ssim of 0.804.

I don't understand why the psnr I tested is not the same as your paper(the psnr of 22.89, the ssim of 0.803 in your paper) , I hope to get your answer.

the pretrained model can't download because the onedrive can't open

Can we use this network for commercial purpose for a small project, as a part of an image processing processes.

pretrained model link is broken

The model download link is broken. Could you kindly check it.

Hi, can you release your unaligned dataset?

Hi, can you release your unaligned dataset?

Hard to put dataset in the right place for just a test.

Is there any hint for put dataset in the right place for at least a test?

The Google Drive link is not found

Hi，
The google drive of pretrained model is not available, could you please check that link?

Best regards

Do 90 real-world training images from Berkeley real dataset have to be cropped to 224x224 When training the model for aligned model?

hey! When i train the model,it reports error like this line 357, in _process_next_batch raise batch.exc_type(batch.exc_msg) KeyError: 'Traceback (most recent call last) . I wonder is this attribute to the data sample size problem?
Actually, The size from three sets for training are

• VOC2012 . 224x224
• testdata_reflection_synthetic_table2 . 224 x224
• real_train . inconsistent ， Mostly larger than 224

Do I need to center-crop 90(actually 89) samples in real dataset to 224x224 to train this model??

Thank you !

About ground-truth of the 90 real training data

Should the real training data be normalized to range [0,1] during training? I find your training code doesn't do that.

Questions about evaluation on Real20

Hi, thanks for your awesome work.

I want to evaluate your model on Real20.

However, the resolution of real20 dataset is too large for prediction.

Which size are you use to evaluate?

GAN loss

Hi，I found that during the training process, the loss of D decreased (maybe to 0.2), and the loss of G increased (maybe to 6). Is this normal?

Slightly different architecture

Hi, I've noticed that the implemented architecture of DRNet is slightly different from the architecture described on the Figure 2 in the paper https://arxiv.org/pdf/1904.00637.pdf:
(1) in the code SELayer consists of Linear-ReLU-Linear-Sigmoid layers, but in the paper convolutional layers are drawn instead of Linear
(2) in the code deconvolution+convolution layers are used after residual blocks, but in the paper only a convolutional layer is drawn.

Regarding synthetic training dataset generation

Hi, The "VOC2012_224_train_png.txt" file lists 15,287 images and the training dataset consists of half of 15,287 i.e. 7,643 images. Is the split for the reflection and transmission layer images random? Can I split the 15,287 images into two equal sets randomly : transmission layer and reflection layer, and go on to combine one image from each set to get 7,643 images containing reflection ?

BrokenPipeError: [Errno 32] Broken pipe

Dear Team,
How to fix the multiple workers on windows? I run that code on windows, and have this error when I set the nThread=8. It's Ok for 0, but the code run too slow and broke down. How to fix it ?

Can we use this network and dataset for commercial purpose?

Is it possible to add license file?
I want to use both code and dataset for my project.

Dataset means this.

We captured a new dataset containing 450 unaligned image pairs that are considerably easier to collect.