chaneyddtt / coarse-to-fine-3d-animal Goto Github PK

@chaneyddtt Hello! Coarse-to-fine-3D-Animal is a nice work. I have a few questions about stage 3 training.

1. Did you try to add supervision on coarse-network output (i.e. SMAL parameters) at stage 3?
2. Will stage 3 supervision only on non-parametric representations worsen predictions of parametric representation?
3. Why feat_multiscale used a detached copy here at stage 3? Shouldn't the graphnet update the multiscale backbone here?

Hi authors,
Is there some API to retrieve the 3D mesh in .ply and .obj format in your method ?

If i have a SMAL model, how can i convert it to 'mesh_down_sampling_4.npz' like you did?

I see you have a reply that mentions coma, and I see a coma folder in your code, but I don't see a save code for np.save(*.npz).

Hi
I can't find files in the project like test_stanford_StanfordExtra_v12.npy, train_animal_pose.npy, train_stanford_StanfordExtra_v12.npy,breeds.csv.
I downloaded Stanford dataset but that doesn't have Stanford extra dataset.

I got the following error while training the stage 3. Note that this error isn't found in stage1 or stage2.

Traceback (most recent call last):
  File "main_meshgraph.py", line 416, in <module>
    main(args)
  File "main_meshgraph.py", line 214, in main
    loss.backward()
  File "/home/napan/anaconda3/envs/c2f/lib/python3.7/site-packages/torch/tensor.py", line 221, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph)
  File "/home/napan/anaconda3/envs/c2f/lib/python3.7/site-packages/torch/autograd/__init__.py", line 132, in backward
    allow_unreachable=True)  # allow_unreachable flag
RuntimeError: function RasterizeFunctionBackward returned an incorrect number of gradients (expected 11, got 10)

Thanks for your excellent work! I'm trying to run the demo, but I got an error 'No such file or directory: 'data/priors/unity_betas.npz'. Will you upload this file later?

the outputs now are generated from the cropped and transformed images, which can not be attached to the original images directly.
what should do if I want to the results aligned on the original images?
replaced the cropped images with original ? it seems not a good idea.
could you give me some instructions please.
many thanks.

@chaneyddtt

Do you still remember the results after stage2 which freezes the coarse network and only trains the subsequent network? Do the results get better before stage 3, or only after stage 3 the results get better?