The code could run successfully, but the output is random. The output of the deep network is good but when doing 'Affine Align', the alignment results between the depth and 3D template are random. The results are different when running the code several times. Sometimes only partial face is reconstructed. I think this may be the depth cannot align well with the template. The 'postprocess.p' is protected so I cannot figure out the problem.

I use GPU and change the run_net.lua file : output = model:cuda():forward(in_im:cuda()):double() but I still got errors:

Running Net...//torch/install/bin/luajit: //torch/install/share/lua/5.1/torch/Tensor.lua:238: attempt to index a nil value stack traceback: //torch/install/share/lua/5.1/torch/Tensor.lua:238: in function 'type' ./torch/install/share/lua/5.1/nngraph/gmodule.lua:233: in function 'recursiveType' **/torch/install/share/lua/5.1/nngraph/gmodule.lua:247: in function 'applyTypeToTable' **/torch/install/share/lua/5.1/nngraph/gmodule.lua:252: in function 'cuda' run_net.lua:27: in main chunk [C]: in function 'dofile' /torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:150: in main chunk [C]: at 0x00405d50 Error using load Unable to read file '//pix2vertex-master/imgs/2327-3-res.mat'. No such file or directory.

Error in run_net (line 13) net_res = load(resname);

Error in runme (line 26) im_pncc = run_net( img, '/**/pix2vertex-master/models/pncc_net.t7', im_file(1:end-4));

Do you know how to fix it?

Hi Matan and Elad,

Thanks for an awesome work! In your paper Fig.9, you show results of each one of your 3 steps results. However, in the code, the non-rigid registration and fine-details reconstruction are packed into a single post-processing function. Is it possible that you provide them as two separate functions?

That would be really helpful in many scenarios; i.e if ground truth coarse shape is given, it's can be feed to your refinement module to do further processing.

Thanks

Hello,

when compiling the code, I found out that the line system(['OMP_NUM_THREADS=1 th run_net.lua -model ' net_path ' -inputs ' matname ' -res ' resname ]); in "run_net.m" doesn't work for me. When typing this: [f,k]=system(['OMP_NUM_THREADS=1 th run_net.lua -model ' net_path ' -inputs ' matname ' -res ' resname ]) I get these results:

k =

127

f =

'/mnt/appl/software/Torch/1.0-goolfc-2018a/bin/luajit: symbol lookup error: /home.others/eb/easybuild/software/Torch/1.0-goolfc-2018a/lib/libTH.so.0: undefined symbol: dgemm_
 '

Does anyone know how to fix this?

Thanks in adavance.

Thanks for you sharing this code, it's quite incredible, however, I cannot open and edit the postprocess.p file, is these any way to open and edit this file.

Any help will be gratefully appreciated!

Hi! First of all, great paper, awesome results! Never had much interest in machine learning, but this paper actually inspired me to learn more and try it out. I've tried to run this implementation, following the instructions in Readme, but I've stumbled an issue when running the code in Matlab (on Ubuntu 16.04 and Matlab R2012a, latest version of Torch). At the postprocessing step I get error:

Error using **repmat**
Too many input arguments.

Error in /home/user/pix2vertex/postprocess.p>postprocess (line 38)

Error in **runme** (line 42)
[vertex, face, texture] = postprocess(img, im_pncc, im_depth, show_figs);

I don't have much experience with Matlab, but according to docs, .p file is a protected file, so I don't really have access and can't really debug what's going on down there. Would be great if you give some pointers!

My best, Ilja

The code runs successfully, but only the front view reconstruction effect can be generated, and the reconstruction of multiple viewpoints in the paper cannot be realized. What can I do to observe the reconstruction model from different viewpoints? like this:

Hello,

I get the following error in some cases (not image specific, encounter it when I run on multiple images)

Starting Detail Extraction
Attempted to access imgh(-0,163); index must be a positive integer or logical.
Error in mesh2fine>ZBuffer (line 635)
Error in mesh2fine>mesoscopic_deformation (line 80)
Error in mesh2fine (line 25)
Error in runme (line 58)
    [ fine_result ] = mesh2fine(pipeline_args); %% 7

Hello all,

I've been experimenting with your code in pytorch, and it is my understanding that there is a final step that is not applied using the python/torch method you've published. The results in this repo seem to be way more accurate than pytorch is able to provide. In that light I have two questions:

1. Is this code runnable using Octave? or alternatively:
2. Taking into account the Matlab version produces much better results. What is the difference between this code and and the pytorch one, does it change any steps for to account for the increase accuracy and result quality, or it misses any final step completely? If yes, which ones, so that I can promptly check that out.

Thank you,

I am glad to check and run this code. I figure out that the final output is somehow not clear and most of the face area is not covered. You can check the attached image. On the other hand in the paper, the same image mesh output is clear and fulllength.,