An unofficial Pytorch extension for occupancy grids and volume rendering in Permuto SDF (CVPR 2023). Original repository: https://github.com/RaduAlexandru/permuto_sdf.

cd py_permuto_sdf
pip install -e .

Requires Pytorch and CUDA. Tested on Pytorch version 1.12.1 + CUDA 11.3 on both Linux and Windows 10.

from py_permuto_sdf import OccupancyGrid, Sphere, RaySampler, VolumeRendering

Most functions and APIs keep the same as the original repo. Instead of the following changes:

# Old
occupancy_grid = OccupancyGrid(256, radius, [0, 0, 0])
# New
occupancy_grid = OccupancyGrid(256, radius, torch.tensor([0, 0, 0], dtype=torch.float, device='cuda'))

# Old
bounding_sphere = Sphere(radius, [0, 0, 0])
# New
bounding_sphere = Sphere(radius, torch.tensor([0, 0, 0], dtype=torch.float, device='cuda'))

Some additional functionalities are extended compared to the original repo, including:

• VolumeRendering.integrate_with_weights now supports not only 3-channel RGB, but also 1-channel depth and 16, 32, 48, 64-channel feature vectors
• New function: OccupancyGrid::update_with_sdf_positions, which supports directly updating the SDF value in the occupancy according to the specified position

Most implementations are derived from the original repo. Thanks for Radu Alexandru's great work.