facebookresearch / DeepSDF

Learning Continuous Signed Distance Functions for Shape Representation
MIT License
1.4k stars 259 forks source link

Test data preparation during reconstruction #83

Open CuiLily opened 3 years ago

CuiLily commented 3 years ago

In my project, I can only get the points on the surface of the object. At present, I have completed the training with ShapeNet's data.

During inference, I made the following changes: I firstly pre-processed ShapeNet's test set (surface samples) to get .ply files. Because these points have 0 SDF value, I just wrote the coordinates of these points and the 0 SDF value of into . npz files. Then I used the new generated .npz file to learn latent code during inference time. But in this way, the reconstructed mesh seemed not similar to ground truth.

I wonder if there is anything unreasonable about what I did?

novauto-nju commented 3 years ago

This is a very clever idea. However, I think this will lose the scalability of the function. Using only zero values to reverse the latent vector will make the function lose the learning of external value expansion