neus-acc with feature grids and hash encoding

[slpmee]

Hi, I am trying to train dtu-scan65 dataset with neus-acc and hash encoding options. the normal and the depth map are not properly estimated. Is there any options that I have to set to the neus-acc? In my opinion, pruning occupancy grid have a bad effect on the training based on hash grid encoding..

This is the command I have tried. ns-train neus-acc --pipeline.model.sdf-field.inside-outside False --pipeline.model.sdf-field.use-grid-feature True --pipeline.model.sdf-field.encoding-type hash --pipeline.model.sdf-field.num-layers 2 --vis tensorboard --data data/sdfstudio-demo-data/dtu-scan65

[niujinshuchong]

@slpmee The mode is sensitive to some hyper parameters beta-init and bias. You could try with ns-train neus-acc --pipeline.model.sdf-field.inside-outside False --pipeline.model.sdf-field.use-grid-feature True --pipeline.model.sdf-field.encoding-type hash --pipeline.model.sdf-field.num-layers 2 --pipeline.model.sdf-field.beta-init 0.3 --pipeline.model.sdf-field.bias 0.5 --vis wandb --trainer.steps-per-eval-image 500 --data data/sdfstudio-demo-data/dtu-scan65