search

microsoft / DIF-Net

Deformed Implicit Field: Modeling 3D Shapes with Learned Dense Correspondence CVPR 2021
MIT License
121 stars 19 forks source link

Training data preparation not working #5

Open blacktime14 opened 3 years ago

blacktime14 commented 3 years ago

Hi, I have an issue regarding data preparation. As you mentioned in Readme, I tried to prepare ShapeNetv2 dataset for training by using mesh_to_sdf. I used get_surface_point_cloud and sample_sdf_near_surface function (like the captured image below), but the result was really different from your uploaded eval dataset. image

Do I need to change detailed arguments for both functions to get the same results?

Selection_779 This is the given data from eval dataset, and Selection_778 This is our failure version.

Thanks in advance.

YuDeng commented 3 years ago

Hi, sorry that the description of the data preparation part is not that clear.

We actually modified the original mesh_to_sdf library to get the data as in the eval dataset:

  1. We normalized the object within a sphere with a radius of 1/1.03 instead of 1 as in the original mesh_to_sdf.
  2. We directly sampled 500k points on the object surface as well as their normals instead of sampling points "near the surface".
  3. We uniformly sampled 500k points within the cube of [-1,1]^3 instead of in the unit spherical space.
  4. For all sampling points, we compute their sdf using the "scan" and "depth" method in the mesh_to_sdf library.

For more details, you can check the supplementary materials of our original paper. Hope it answers your question.