Questions about Evaluation Pipeline and Metrics

[colinzhenli]

Great work! I have a question regarding the evaluation metrics, such as the Chamfer Distance and F-score.

• In baseline methods like SA-ConNet and ConvOcc, meshes were reconstructed from SDF using the marching cubes algorithm. Subsequently, a point cloud was sampled from the predicted mesh to compute the CD and F-score.

• While in your code, I noticed that you reconstructed a dense point cloud using a gradient-based algorithm similar to NDF's approach. Then, you directly computed the CD and F-score between the predicted point cloud and the ground truth point cloud.

• You also mentioned using marching cubes for UDF to visualize the meshes.

My question is: In your published quantitative figures, did you use the SDF-to-mesh-to-point cloud and UDF-to-point cloud methods, or did you reimplement the same evaluation pipeline?