⭐ 3inFusion

[Work in progress] Showcase project for the thr3ed_atom package. We use a diffusion network to model semantic variations of a given single 3D scene which is self similar and stochastic. Following is a teaser of this concept: The diffusion model is trained on fixed sized 3D crops of the original 3D voxel-grid (containing the object or the scene). At test-time, samples of original resolution are drawn using the trained model. This idea essentially replaces the GAN in 3inGAN with DIFFUSION. The code can be located under the thre3d_elements/thre3infusion directory.

Some results in Action:

G.T. 3D scene:

rendered_video.mp4

Random samples:

samples_75000.mp4

relu_fields

Official implementation of the paper: ReLU Fields: The Little Non-linearity That Could. The implementation uses the PyTorch framework.

Mitsuba 3 tutorial

Check out this Mitsuba 3 inverse rendering tutorial based on ReLU Fields.

Radiance Fields Experiments

Data:

The data used for these experiments is available at this drive link. Download the synthetic_radiance_fields_dataset.zip and extract it in some folder on your disk. For easy access create a /data folder in this repo-root and extract the zip in something like /data/radiance_fields.

Run the optimization:

Running the optimization is super simple 😄. Just use the script train_sh_based_voxel_grid_with_posed_images.py python script with appropriate cmd-args. Following example runs the training on the Hotdog scene:

(relu_fields_env) user@machine:<repo_path>$ python train_sh_based_voxel_grid_with_posed_images.py -d data/radiance_fields/hotdog -o logs/rf/hotdog/

Render trained model:

Use the render_sh_based_voxel_grid.py script for creating a rotating/spiral 3D render of the trained models. Following are all the options that you can tweak in this render_script

Usage: render_sh_based_voxel_grid.py [OPTIONS]

Options:
  -i, --model_path FILE           path to the trained (reconstructed) model
                                  [required]
  -o, --output_path DIRECTORY     path for saving rendered output  [required]
  --overridden_num_samples_per_ray INTEGER RANGE
                                  overridden (increased) num_samples_per_ray
                                  for beautiful renders :)  [x>=1]
  --render_scale_factor FLOAT     overridden (increased) resolution (again :D)
                                  for beautiful renders :)
  --camera_path [thre360|spiral]  which camera path to use for rendering the
                                  animation
  --camera_pitch FLOAT            pitch-angle value for the camera for 360
                                  path animation
  --num_frames INTEGER RANGE      number of frames in the video  [x>=1]
  --vertical_camera_height FLOAT  height at which the camera spiralling will
                                  happen
  --num_spiral_rounds INTEGER RANGE
                                  number of rounds made while transitioning
                                  between spiral radii  [x>=1]
  --fps INTEGER RANGE             frames per second of the video  [x>=1]
  --help                          Show this message and exit.

Coming soon ...

The Geometry (occupancy), Pixel Fields (Images) and the Real-scene experiments will be setup in the thre3d_atom package soon.

BibTex:

@article{Karnewar2022ReLUFields,
    author    = {Karnewar, Animesh and Ritschel, Tobias and Wang, Oliver and Mitra, Niloy J.},
    title     = {ReLU Fields: The Little Non-linearity That Could},
    journal   = {Transactions on Graphics (Proceedings of SIGGRAPH),
    volume    = {41},
    number    = {4},
    year      = {2022},
    month     = july,
    pages     = {13:1--13:8},
    doi       = {10.1145/3528233.3530707},
}

Thanks

As always,
please feel free to open PRs/issues/suggestions here 😄.

cheers 🍻!
@akanimax 😎 🤖