search

NeuralCarver / Michelangelo

[NeurIPS 2023] Michelangelo: Conditional 3D Shape Generation based on Shape-Image-Text Aligned Latent Representation
https://neuralcarver.github.io/michelangelo/
GNU General Public License v3.0
341 stars 11 forks source link

How to get the BCE Loss? #10

Closed XiongFenghhh closed 5 months ago

XiongFenghhh commented 5 months ago

Thank you for your great work!

When training decoder, you used BCE Loss. But how could we get the ground truth occupancy of the query point?

Maikouuu commented 5 months ago

Thank you for your interest.

We utilize the SDF function in the pysdf library to compute the SDF of each point.

XiongFenghhh commented 4 months ago

Thank you for your interest.

We utilize the SDF function in the pysdf library to compute the SDF of each point.

Hello, I tried to reproduce training procedure recently, but didn't get a good result. Could you give me some advice?

  1. How many points should be set for the encoder's input point cloud during training?
  2. What is the sampling strategy for the query points during training? Is it sampling within an NxNxN volume and near the mesh surface? If volume sampling is used, what is a suitable value for N?

Thanks!

Maikouuu commented 3 months ago

Hi,

  1. The encoder uses 4096 surface points as input;
  2. The ratio of volume points to near points (points close to the mesh surface) is 1:1.

Hope this information is helpful to you.