Could you provide a set of test data, such as the visualized data in the paper?

[Azzurro96]

A working example would be helpful to understand the code.

[cedricpradalier]

Could you provide just one commande line to see how to get started? I understood that a PCD needs to be converted in npz first, but the path to get an SDF and potentially a mesh is a bit confusing in the code so far.
Thanks

[Azzurro96]

Could you provide just one commande line to see how to get started? I understood that a PCD needs to be converted in npz first, but the path to get an SDF and potentially a mesh is a bit confusing in the code so far. Thanks

@cedricpradalier I can't get it work either. The point cloud data needs to be converted to npz fitst. I changed the input file path to make it run without error. But it seems that a conf file is necessary. I tried to write a conf file with the information in the paper. But my conf file can not help me get the same result shown in the paper. Did you make any progressing so far?

[Ouasfi]

Could you provide just one commande line to see how to get started? I understood that a PCD needs to be converted in npz first, but the path to get an SDF and potentially a mesh is a bit confusing in the code so far. Thanks

@cedricpradalier I can't get it work either. The point cloud data needs to be converted to npz fitst. I changed the input file path to make it run without error. But it seems that a conf file is necessary. I tried to write a conf file with the information in the paper. But my conf file can not help me get the same result shown in the paper. Did you make any progressing so far?

Hi guys,

Sorry for the delayed response. I added detailed instruction about the data processing and the training in addition to a working example that you can run with:
python train_socc.py --shapepath data/Faust/real_11/ --device 0 --exp_dir experiments/Faust/real_11/ --n_points 1024 --sigma 0.0 --n_surface 1000 --lamda_max 10 --n_minimax 1000

To select the best checkpoint using the input pointcloud run:

python eval.py --device 0 --shapename data/Faust/real_11/ --results_dir experiments/Faust/real_11/

Please let me know if this works for you. You should get a chamfer-L1 of 0.002437 and a normal consitency of 0.96.

[cedricpradalier]

Thanks, I could install and run your examples. The Chamfer distance is consistent with your values.
I would suggest adding a "requirements.txt" file. I add to install the following packages:

 2015  pip install --upgrade  pymcubes==0.1.4
 2017  pip install --upgrade  pyhocon
 2019  pip install --upgrade  wandb
 2021  pip install --upgrade  napf

With Python 3.8.16, the specific version of PyMCubes was required.
Also, with scipy==1.10.1, I had to fix the renaming of the parameter n_jobs to workers in the kdtree.
And I had to move compute_dists_flann before its use as default argument to validate_mesh.
Here is the diff
SparseOcc.txt

Now a practical question: how can I output the generated mesh when evaluating a model ?

Thanks

[cedricpradalier]

Actually, I answered my own question by modifying eval.py:

    pred_mesh_o3d = o3d.geometry.TriangleMesh( o3d.utility.Vector3dVector( mesh.vertices),
                              o3d.utility.Vector3iVector( mesh.faces) )
    o3d.io.write_triangle_mesh("mesh.ply",pred_mesh_o3d)

[cedricpradalier]

Another request if I may: even though I read the paper, I don't see exactly how I can play with the parameters to change the resolution of the reconstruction. For context, I have pointclouds describing 10m long object with 5 million points, and I'm interested in a model that would keep 2mm-scale details. Can you add some documentation of the parameters in train and eval, possibly by linking them to the variables in the paper ? Thanks

[Star-Explorer88]

Hello author, thank you for your contribution. May I ask how to visualize the evaluation grid? Do I need to add scripts

[Ouasfi]

Hi @Star-Explorer88, could you clarify what you mean by "evaluation grid"? In this context, the evaluation begins with a 3D grid of points represented as a tensor of size (resolution, resolution, resolution,3). These points are passed through the occupancy model, which produces an output grid of size (resolution, resolution, resolution) representing the occupancy values. The final mesh is then extracted from this grid using the Marching Cubes algorithm. Then this mesh is compared the ground-truth mesh to compute the evaluation metrics.