Thanks for the wonderful work~! I've done the separate training of two sub-network(depth prediction + texture mapping) with given data, the result looks good. But there's no code for joint training in origin paper(4.2), is it possible to add loss function to infer.py ?

Hi! For training WC flag augmentations=True is passed in dataloater https://github.com/cvlab-stonybrook/DewarpNet/blob/bdf01e613792f940c6ec45c4e3ebc5789b1703c6/trainwc.py#L40

So, there is augtexnames.txt file: https://github.com/cvlab-stonybrook/DewarpNet/blob/bdf01e613792f940c6ec45c4e3ebc5789b1703c6/loaders/doc3dwc_loader.py#L41

I couldn't find it in repository. How should this file look like?

Thanks~

Hi! I have been using this repo for months and have trained many models. For bm training, I have trained many models but none of them are usable. I noticed that for loss of bm model, there are comment session of SSIM loss. Should we add this to bm training? What is the correct setup for the loss of bm training? I tried 10* l1loss + 0.5*rloss but the performances are terrible.

Hello, I'm confused of the implementation of dense block https://github.com/cvlab-stonybrook/DewarpNet/blob/f91bdf6248789e58c836cee828493d962396ec0c/models/densenetccnl.py#L62-L66 1、while i >0, even no use the layer 2、do you want use sum instead of torch.cat?

Hi, I'm a student studying your project, DewarpNet! First of all, Thank you for sharing this awesome project :)

I read a paper, and Refinement Network to adjust for illumination effects in the rectified image is obviously recorded in the paper. But, I can't find this network in this code.

Is there any training code or pretrained model?

I implemented the two-phase training schedule by exploiting the provided doc3D dataset (~10w). However, the output provided by infer.py is not correct and it seems that the networks are not properly trained.

To pinpoint the problem, I download the provided official pre-trained models. With official wc network + official bm network, the results are pretty good. With my trained wc network + official bm network, the results look fine. With my bm network + official wc network, the result is not correct, and it is the problem.

The loss of training bm seems fine (with lr = 1e-4, the train and val MSE loss decreases continuously and approximate 5 * 1e-5 at epoch 50). The mentioned solutions in other issues like pytorch version (pt1.4 + pt0.4.1) are tried but do not help.

The only modification is in doc3dbmnoimgc_loader.py, line 29, the self.altroot = "xxx/swat3d". Since I can not find this folder in the project, it is altered to self.root according to my understanding.

Would you please point out other possible reasons?

Looking forward for your reply, thanks!

Hello, Thanks for sharing your code.

I was able to run the model but it seems doesn't perform as well as illustrated by the paper.

First, I tried the provided test images under /eval/inp, running with infer.py as suggested by README.md, the output doesn't look as good as those in /eval/uw;

Second, I tested inference with a few photo shots of an article angled on a black background, no other objects were in the photo. However, the output is either highly twisted or not fully angle-corrected.

Other info: Since there is no requirements info provided. I'm running with pytorch==1.3, CPU only, I also replaced numpy.misc.imread function to PIL and fixed dtypes.

Thank you, Julian

Hi, I find that the output world coordinate is resized to (128, 128) frim (256, 256). So for joint learning, which size should the supervision put on, the size (256, 256) or (128, 128)?

Thank you!

Hi，Thanks for your released code. I have one questions：

For shape network, whether the intensity and color of each training image are also randomly jittered. In code,

1. For function color_jitter(im, brightness=0, contrast=0, saturation=0, hue=0): , should we use the parameter "saturation=0, hue=0";
2. Should we use the function change_intensity()， I find you comment out this code.

What is your implementation of these settings in your public model？

Thanks for your solid work!

Could you provide the rectified images of the refinement network or the pre-trained model? We want to compare the results of our implementation.

Thank you!

Please update a requirement list here, cuz a error appears when I run infer.py and error is the 'scipy.misc' module doesn't have 'imread' and others, and I think the reason may be the version mismatch. Thx!!

Hi, I'm working on a project that use text in image with extracting OCR machine and your source code is great and can increase accuracy and if I use your source be sure I will consider citing your work. But the problem is, how can I do that ? for example How can I write pip install ...??? Best regard,

Hi @sagniklp,

I was wondering if you have any details about the camera parameters/model used when rendering doc3D in Blender? Is this constant throughout the dataset or do the camera parameters (other than position & orientation) change? Do you know the position of the camera in the scene relative to the meshes for each sample from doc3D?

Hello!Thanks your great work! I used the final models you provided,and run infer.py to dewarp the DocUNet images, but got the MS-SSIM:0.449739 ,LD: 9.131016 with MATLAB R2018b and pytorch 0.4.1

I've completed the form a few days ago, could you please send the username and password to [email protected]? Thanks so much for your work!