Question about evaluation speed

[add2333]

Hi,

Thanks for the great work! I have a question about evaluation speed.

I'm evaluating on scenes with 500+ views (e.g., NRGBD, kf=2). I find that the accuracy/completion computation runs KDTree queries on the full point cloud, which is extremely slow. This makes large-scale evaluation impractical (~ 1 hour per scene on a100 40g & Xeon(R) Gold 6348 CPU) .

A few questions:

1. What evaluation speed do you typically see?
2. Should I set --voxel_size to downsample before metric computation? It's in the code but defaults to 0.0 and isn't used in any example scripts.
3. Or is there something I'm missing?

Thanks!