autonomousvision / sdfstudio

A Unified Framework for Surface Reconstruction
Apache License 2.0
1.9k stars 179 forks source link

custom image Train works fine in nerfstudio, but not in sdfstudio #288

Closed hanjoonwon closed 3 months ago

hanjoonwon commented 5 months ago

https://www.notion.so/sangyul/cb39a9b461e44505bd74058e353dd406?pvs=4#696214f99dcd477fb2c113d38d570fa0 in nerfstudio train well..

However, when I try to run sdfstudio's train in wandb, it doesn't work and the result is not good. https://wandb.ai/ju805604/sdfstudio/runs/5h16cbww?workspace=user-ju805604

i trained with ns-train neus-facto --pipeline.datamanager.train-num-rays-per-batch 1024
--pipeline.model.sdf-field.use-grid-feature False --pipeline.model.background-model mlp
--pipeline.model.mono-depth-loss-mult 0.0 --pipeline.model.mono-normal-loss-mult 0.0 --vis wandb --pipeline.model.sdf-field.inside-outside False --trainer.steps_per_save 5000
--trainer.steps-per-eval-image 5000 --trainer.max-num-iterations 300000 --experiment-name nolossowl sdfstudio-data --data /workspace/processdata/owl2 --include_mono_prior True

I tried disabling loss and leaving it at 0.1 or 0.05, but it's still not good.

my datasest is here https://drive.google.com/drive/folders/1lHsiW8MGQcVTrQT9LtJjH8Z6ChMiiVlf?usp=sharing