search

HTDerekLiu / neuralSubdiv

prototype code for neural subdivision
Mozilla Public License 2.0
115 stars 14 forks source link

tosca shape #11

Open therateful opened 2 months ago

therateful commented 2 months ago

Hi, I am trying to subdivide the shapes in the Tosca dataset, as shown in Table 2 in the paper, but I found that many of the shapes have boundaries, how did you subdivide these shapes and get the results?

therateful commented 2 months ago

In addition, the Metro tool outputs two mean distances, which is used as the average surface distance

HTDerekLiu commented 2 months ago

Re: TOSCA I closed the mesh (see attached). Alternatively, you could make it work by doing some padding too. tosca_closed.zip

Re: Metro Sorry it has been a while, I couldn't recall what are the numbers and which one I used.

therateful commented 2 months ago

thank you very much

therateful commented 2 months ago

Re: TOSCA I closed the mesh (see attached). Alternatively, you could make it work by doing some padding too. tosca_closed.zip

Re: Metro Sorry it has been a while, I couldn't recall what are the numbers and which one I used.

I'm sorry, but I might have one more question for you. I performed mesh simplification on the original meshes you shared, then subdivided them with neural subdivision, and finally quantitative analysis with Metro tools, but got poor results on some models. and I have observed that the simplified mesh model is a little bad, so could you please share your simplified mesh?

therateful commented 2 months ago

Or am I using the Metro tool incorrectly? I used the simplest assessment, "Metro A.obj B.obj."

HTDerekLiu commented 2 months ago

Based on the file name, this should be the ones I was using (not 100% sure, it has been a while)

therateful commented 2 months ago

Based on the file name, this should be the ones I was using (not 100% sure, it has been a while)

Thanks for sharing, but it seems to require additional permissions to access this connection

therateful commented 2 months ago

Based on the file name, this should be the ones I was using (not 100% sure, it has been a while)

I used the coarse mesh you provided and subdivided it by neural subdivision, but the Hausdorff distance calculated by Metro between the prediction model and the ground-truth model of cat is still 134.xxxx >>>2.08 (the figure in paper ). However, when I work with other models, non-TOSCA models, the values are normal.Is there a problem with the model of the Tosca dataset？Or am I using the Metro tool incorrectly?

HTDerekLiu commented 2 months ago

I guess the mesh scale may be incorrect? Here are some pre-trained weights on centaur and subdivision results for you to debug. https://drive.google.com/file/d/197Nm5Z6uI6uHmf-xWXWH0ljXrUuwobzk/view?usp=sharing

therateful commented 2 months ago

I guess the mesh scale may be incorrect? Here are some pre-trained weights on centaur and subdivision results for you to debug. https://drive.google.com/file/d/197Nm5Z6uI6uHmf-xWXWH0ljXrUuwobzk/view?usp=sharing

You are right, it is a matter of scale, but how to adjust the scale correctly?

therateful commented 2 months ago

I guess the mesh scale may be incorrect? Here are some pre-trained weights on centaur and subdivision results for you to debug. https://drive.google.com/file/d/197Nm5Z6uI6uHmf-xWXWH0ljXrUuwobzk/view?usp=sharing

And I'm a little confused, why is there such a big difference between the scale of subdivision output mesh and the ground-truth mesh？

HTDerekLiu commented 2 months ago

The method is not scale invariant so we normalize the mesh to a unit box. In the previous data, *_s0.obj, *_s1.obj, *_s2.obj are the subdivision results. *_target.obj are the target ground truth shapes used to compute losses.

therateful commented 2 months ago

Ok, I see. Thank you for your help.