Skip to content

Latest commit

 

History

History
28 lines (24 loc) · 1.87 KB

README.md

File metadata and controls

28 lines (24 loc) · 1.87 KB

Latent-SDF

Intro

  • This is the repository of our final project of Deep Learning, 2023 summer, NYCU.
  • Our goal is to combine two method: Geo-NeuS and Latent-NeRF. We aim to use latent paint method in Latent NeRF, but instead of using mesh and latent texture, we replace the mesh with the SDF network obtained from Geo-NeuS, and replace the latent texture with trainable latent radiance network.
  • This project is based on Geo-NeuS and Latent-NeRF. Part of the code is from Textured-NeuS.

Environment

  • Environment need to satisfy the requirement in Geo-NeuS and Latent-NeRF, please refer to these 2 repo to build your environment.

Usage

  1. train the model from Geo-NeuS, and put the model in neus_ckpt
    • change the model path in confs/train_config.py, neus_ckpt_path
  2. make sure the dataset of Geo-NeuS exist
    • change the dataset path in confs/womask.conf
    • change the case name in train_config.py, case
  3. Execute: python Trainer.py --log.exp_name "exp_name" --guide.text "guide text", you will see the result in experiments/exp_name/

Result

  • Although we do not get better result as Latent-NeRF, but our network's storage requirement is relatively small(3.8 MB), which is far smaller than regular mesh.

Future work

  • Complete the README
  • GeoNeusTrainer compability
  • Complete pixel-based training(using Deepfloyd IF model)
  • Use different differentiable SDF rendering method
  • Filter spurious region in SDF networks
  • Finding optimal network architectures and hyperparmeters
  • Also train background NeRF(idea from NeRF++)