akanazawa / cmr

Project repo for Learning Category-Specific Mesh Reconstruction from Image Collections
MIT License
474 stars 84 forks source link

Update to PyTorch1.x and Python3, remove chainer/cupy dependancy. #24

Closed chenyuntc closed 4 years ago

chenyuntc commented 4 years ago

Nice work! I've been working on a similar task, your code is clean and intuitive, a good start for the project.

But I struggle to set up the env. So I make some updates:

The repo is in chenyuntc/cmr. I can make a pull request if you would like.

shubhtuls commented 4 years ago

Hi @chenyuntc , Thanks for making the code python3 and pytorch 1.x compatible! I think given Angjoo and I are no longer actively maintaining this codebase, will will just link to your implementation from our README to direct interested people.

Although would it be possible for you to share some sample results that the updated code gives so we're sure that the updates to the neural renderer etc. didn't break anything? Thanks!

chenyuntc commented 4 years ago

Hi @shubhtuls

Following is the comparison of evaluation results and visualization:

Results from the pretrain model (500 epoch)

results.mat mean iou 0.706, pck.1 0.815, pck.15 0.931
results_sfm_camera.mat mean iou 0.741, pck.1 0.923, pck.15 0.983
results_sfm_meanshape_sfm_camera.mat mean iou 0.589, pck.1 0.913, pck.15 0.975

Results of the model trained from scratch (200 epoch)

results.mat mean iou 0.693, pck.1 0.798, pck.15 0.923
results_sfm_camera.mat mean iou 0.737, pck.1 0.928, pck.15 0.98
results_sfm_meanshape.mat mean iou 0.565, pck.1 0.783, pck.15 0.914

image1

image2

I will update the results after it finishes training.

shubhtuls commented 4 years ago

Thanks a lot @chenyuntc for the efforts in updating the code and sharing sample results!

chenyuntc commented 4 years ago

Hi @shubhtuls, after training the model from scratch with my code for 500 epochs, the results are the following.

  1. IoU curve is very close to the figure in the paper.

image

  1. Interactive mesh visualization please check my mesh visualization in plotly. Plotly only supports face color, no UV map :cry:

I think the issue is almost done. Feel free to close it.