search

CyberAgentAILab / SuperNormal

[CVPR 2024] Official implementation of "SuperNormal: Neural Surface Reconstruction via Multi-View Normal Integration"
MIT License
127 stars 6 forks source link

High memory consumption #6

Open MickShen7558 opened 4 months ago

MickShen7558 commented 4 months ago

Hi,

Thank you for the work and the code!

When I tried to run the DiLiGenT-MV benchmark objects, it seems that the final saving requires very large memory:

129deddcfea7e2e8140b56f12d5c1c2

So is there a way to sacrifice the performance by a little bit and get the result?

xucao-42 commented 4 months ago

Have you installed pyembree library? This step is to find the ray-mesh intersection for geometry evaluation.

Or you can bypass this step by setting val.eval_metric_freq as a larger number than train.end_iter in the configure file so that the evaluation will not be performed. To still extract the mesh, you can set val.val_mesh_freq as 5000, for example, so that a mesh will be extracted for every 5000 step.