Open hskAlena opened 1 year ago
Hi, I think the current implementation of neus-acc has some problem with occupancy grid updates. The update schedule is hard coded and chosen based on one of the dtu dataset scene, so it might not work well on other dataset. Will check later.
I ran NeuS-acc with blender lego dataset. But it does not output any recognizable mesh or rgb results. Is there anything wrong in my command line?
I ran with the following code below.
ns-train neus-acc --pipeline.model.sdf-field.inside-outside False \ --vis tensorboard --pipeline.model.background_color="white" \ --pipeline.model.near_plane 2.0 --pipeline.model.far_plane 6.0 \ --pipeline.model.overwrite_near_far_plane True \ --trainer.steps_per_eval_image 2000 --pipeline.model.background_model none \ --trainer.steps_per_save 4000 --trainer.steps_per_eval_batch 2000 \ --experiment-name neus-acc-beta0.1-bias0.8-lego \ blender-data --data data/blender/lego
And the results show like this.
I guess this error is related to this TODO, do you have any code fix suggestion in this line? https://github.com/autonomousvision/sdfstudio/blob/b511d9e8568e680b44844264e81ff69b632bb4c7/nerfstudio/model_components/ray_samplers.py#L1391
Thanks a lot!