README.md

Chapter 6: Representing Geometry and Appearance with Neural Networks

This chapter compares voxel grids and point clouds with learned occupancy, signed-distance, and appearance fields. The examples share helpers in this folder, but do not need code or outputs from another chapter.

Setup

Use Python 3.12. From Chapter06, install the packages for the CPU examples:

python -m pip install numpy scipy scikit-image matplotlib

The occupancy, SDF, and appearance examples also need a Vulkan-capable GPU and SlangPy:

python -m pip install slangpy==0.42.0

Pick an example

Run each command inside the folder listed, so figures and checkpoints go to the right place. The linked READMEs explain the individual experiments.

Folder	Run	Covers
3d_representations	python run_section3.py	Voxelization, point clouds, and NumPy neural networks; CPU only.
neural_occupancy_fields	python run_section.py	Learning inside/outside tests and extracting a mesh.
neural_signed_distance	python run_section.py	Learning SDFs, sphere tracing, and constructive solid geometry.
appearance_fields	python run_section.py	Albedo, view-dependent color, and a feature-grid/MLP comparison.

For example, starting from Chapter06:

cd 3d_representations
python run_section3.py

These drivers save plots in their section's figures/ folder. The appearance example needs appearance_fields/checkpoints/ckpt_sdf_supervised.npz, which is included. If it is missing, run python train_one.py supervised 1500 inside neural_signed_distance, then copy the resulting checkpoint into appearance_fields/checkpoints/.

Book figures

Files named fig6_N_*.png match the book's Figure 6.N. Files named extra_*.png are supplementary plots that do not appear in the book.

Book figure	Output file	Produced by
6.1 Positional encoding bands	neural_occupancy_fields/figures/fig6_1_encoding_bands.png	run_section.py or run_figures.py
6.2 Sphere tracing the neural SDF	neural_signed_distance/figures/fig6_2_spheretrace.png	run_section.py or run_figures.py
6.3 An SDF from surface points alone	neural_signed_distance/figures/fig6_3_points_alone.png	run_section.py or run_figures.py
6.4 CSG on distance fields	neural_signed_distance/figures/fig6_4_csg.png	run_section.py or run_figures.py
6.5 Albedo field	appearance_fields/figures/fig6_5_albedo.png	run_section.py or run_figures.py
6.6 View dependence	appearance_fields/figures/fig6_6_viewdep.png	run_section.py or run_figures.py
6.7 Pure MLP vs feature grid	appearance_fields/figures/fig6_7_hybrid.png	run_section.py or run_figures.py
6.8 The Stanford bunny	figures/fig6_8_bunny.png	make_bunny_comparison.py

The optional bunny examples use the bundled bunny.ply and additionally need trimesh and embreex. On machines with multiple GPUs, set CHAPTER6_GPU to a vendor or adapter-name substring, such as nvidia; see the occupancy section's README for details.