This is a (hopefully) well-commented implementation of Plenoxels in PyTorch (more comments are still being added to improve explainability). The relevant paper is Plenoxels: Radiance Fields without Neural Networks.

The theoretical background is explained in a series of posts below. Additionally, the incremental build-up of the code will hopefully aid in a more ground-up understanding.

• Part 1
• Part 2
• Part 3
• Part 4
• Part 5
• Part 6 (Work In Progress)

Currently, the code is set to reconstruct a cube from 13 views of it. These views are stored in images/cube/training. The corresponding camera positions are in cube_training_positions(). Simply running plenoxels.py will do the training, and store the reconstructed world in reconstructed.pt. It will also show the reconstructed views from the original training viewpoints in the images/frames directory. The renderings during the training (which will be stochastic samples) are stored in the images/reconstruction directory.

Note that for purposes of speed, the voxel grid is restricted to 40x40x40 for the cube, and 40x40x60 for the table.

Training

Reconstruction

Partial Reconstruction of a Table from a Single Image

Training Image of a Table

Note: The table image above is taken from the Amazon Berkeley Objects Dataset. The ABO Dataset is made available under the Creative Commons Attribution-NonCommercial 4.0 International Public License (CC BY-NC 4.0).