search

vishwa91 / MINER

MINER: Multiscale Implicit Neural Representation
MIT License
43 stars 10 forks source link

Scales and Resolution #4

Open csyhping opened 3 months ago

csyhping commented 3 months ago

Hi @vishwa91 , thanks for your great work. I have one question about the 'four scales.' In the paper, you've mentioned four scales, from coarse to fine, but there seems to be no specific description of the resolution of each scale. I think it should be something like LOD. May I ask the exact resolution for each scale for 3D point cloud fitting? Thanks!

csyhping commented 3 months ago

@vishwa91 ， in the code L27, there is a scale = 0.5, how should I understand this ?

vishwa91 commented 3 months ago

@csyhping thank you for your interest in our work.

  1. scale = 0.5 : this downsamples the target volume. Often times, loading a volume with 1024x1024x1024 voxels is infeasible, so I run it at lower resolution to get a quick result. All the visualizations in the paper were with scale=1, while some of the ablations and sweeps were with scale=0.5.
  2. Number of scales being four means the voxels were 1024x1024x1024, 512x512x512, 256x256x256, and 128x128x128. However, note that since INRs are continuous in nature, these resolutions are somewhat an arbitrary choice.

Hope that answers your questions!