Scale of sdf? about dualoctreegnn

[XinYu-Andy]

Also, I am curious about why we need scale the mesh into [-0.5,0.5], since i noticed that when you sample the pointcloud from the mesh, the points are then scaled back into [-1,1]?

[wang-ps]

In my initial experiments, I did scale the mesh into [-1, 1]. And the SDF is calculated on such a scale.

And since I reuse the evaluation code and results of ConvONet, ConvONet scales the mesh into [-0.5, 0.5], and I also scale the input point cloud and output mesh into [-0.5, 0.5] while keeping SDF fixed.

When training the network, the point cloud in [-0.5, 0.5] is rescale to [-1, 1] ( ref link )

[XinYu-Andy]

Thanks for your reply!

[XinYu-Andy]

Hi,

Sorry for bother, I have another question.
I am wondering whether I only need to change the "size=128" to "size=256" in the "tools/shapenet.py" for generating SDF with voxel size=256 and the corresponding "point cloud", "points", and "occupancy"?
Are there any other code I need to modify?

Thanks!

[XinYu-Andy]

Because I find it will not work when I train an implicit autoencoder model with SDF resolution=256, while it works well with SDF resolution=128.

[XinYu-Andy]

Because I find it will not work when I train an implicit autoencoder model with SDF resolution=256, while it works well with SDF resolution=128.

I suspect there are some inconsistencies in the function "def sample_sdf()".

[wang-ps]

The code can work with depth=8 since I have tried it in my personal experiments before.
I suggest understanding the code and checking all the places related to depth, including the code and the *.yaml configuration files.

[XinYu-Andy]

Thanks for you reply.

In fact, I am not concerning about depth, but the size (i.e., resolution of SDF).
In fact, I have checked the code after I opened this issue, and find that it seems some codes should be modified, e.g., this line and this line.
The (xyz<127) should be changed into (xyz<size-1), while the (xyzs/64-1) should be changed into (2*xyzs/size-1).

[wang-ps]

Good job! Yes, you are right!

[XinYu-Andy]

Thanks for your patient, I am new in this field, I still have another trivial question :{
Is there a way to directlt compute the SDF value near the surface? I understand that in your codebase, you firstly compute a SDF voxel, then do interpolation to get SDF value for query points near the surface.
This may poses several issues:

1. SDF might be not accurate.
2. Pre-computing sdf voxel is costly or useless since many of them will be not used.
   I am seeking a way to compute sdf for irregular points efficiently.

[XinYu-Andy]

Thanks for your patient, I am new in this field, I still have another trivial question :{ Is there a way to directlt compute the SDF value near the surface? I understand that in your codebase, you firstly compute a SDF voxel, then do interpolation to get SDF value for query points near the surface. This may poses several issues:

1. SDF might be not accurate.
2. Pre-computing sdf voxel is costly or useless since many of them will be not used.
   I am seeking a way to compute sdf for irregular points efficiently.

Also, why do we set the mesh scale to 0.8? I do think if we set it more close to 1 (e.g., 0.9), the produced sdf could recounstruct the mesh better because the sampled points are uniform grids in [-1,1]^3. What are the considerations when setting this value (i.e., mesh_scale)?

[wang-ps]

Yes, it is OK to set it to 0.9.
The value 0.8 is relatively conservative. Since the dataset construction process is time-consuming, I did not try it in my experiment after I first constructed the dataset with the mesh scale factor as 0.8.

Also, why do we set the mesh scale to 0.8? I do think if we set it more close to 1 (e.g., 0.9), the produced sdf could recounstruct the mesh better because the sampled points are uniform grids in [-1,1]^3. What are the considerations when setting this value (i.e., mesh_scale)?

[wang-ps]

SDF might be not accurate

Yes, there do exist approximation errors with a resolution of 128^3 (less than 1% of the bounding box) .

Pre-computing sdf voxel is costly or useless since many of them will be not used. I am seeking a way to compute sdf for irregular points efficiently.

The OpenVDB library contains such functions. However, I am not familiar with OpenVDB either and I could not provide more information about it.

[XinYu-Andy]

Thank you very much!