Different range for sampling coordinates and network predictions

[lykius]

Hi @jchibane, Thanks for sharing your work (and nice code btw).

I have some doubts on the point cloud generation algorithm in ndf/models/generation.py. More precisely:

• samples are created in the range [-1.5, 1.5] here; if I understand it correctly, this is because pytorch grid_sample requires the grid coordinates to be in the range [-1, 1] and you add and extra 0.5 because the network was trained with coordinates possibly out of the [-1, 1] range
• df_pred though should be distances predicted by the network for coordinates in the range [-0.5, 0.5] since the udf ground truth is computed here on coordinates belonging to the range [-0.5, 0.5]
• thus, I think that the computation samples = samples - F.normalize(gradient, dim=2) * df_pred.reshape(-1, 1) here is moving samples in the range [-1.5, 1.5] with udf predictions for the range [-0.5, 0.5]

Also: since samples x and z dimensions are not swapped before feeding them to the network (as it is done for training), it seems to me that samples and udf predictions are in a different reference system.

What am I missing? Thanks in advance!

[jchibane]

@lykius thanks! to your points:

• Yes, exactly - for the grid_sample function we need to convert to [-1,1] , additional 0.5 serves as slack - not strongly necessary theoretically.

• Points in [-1,1] in grid sample coordiantes correspond to [-0.5,0.5] in mesh coordinates, so there is no mismatch.

• The movements are predicted in mesh space, the grid_sample space is just intermediate and only important for implementation. The network is supervised with mesh space and the results are interpreted in mesh space - so this is correct as is.

• We are sampling random points in a symmetric cube - so we don't need to swap any coordinates.

Best Julian

[Wi-sc]

Hi @jchibane,

I think @lykius is right here. The code here should be samples = samples - F.normalize(gradient, dim=2) * df_pred.reshape(-1, 1)*2. Because your inputs are grid coordinate points [-1,1] and your outputs are udf under mesh space [-0.5,0.5]. Here the samples are grid coordinate points instead of mesh space points. You move the grid coordinates points instead of mesh space points here.