How city-super/gsdf works
Project type and purpose
• GSDF is a dual‐branch neural rendering and reconstruction framework combining 3D Gaussian Splatting (“GS branch” under gaussian_splatting/…) and neural Signed Distance Fields (“SDF branch” under instant_nsr/…).
• Its goal is joint training of two complementary scene representations, with mutual guidance to improve both rendering quality and geometric reconstruction.
File structure
• Top‐level scripts:
• train.py/render.py/metrics.py/launch.py orchestrate training, rendering, and evaluation.
• configs/… contains YAML configs for datasets and experiments.
• Core packages:
• gaussian_splatting/… implements the GS branch: argument parsing, gaussian renderer, scene loader, LPIPS module, and two C++/CUDA submodules (diff-gaussian-rasterization and simple-knn).
• instant_nsr/… implements the SDF branch: dataset loaders, model definitions (NeRF/SDF), training systems (criterions, optimizers), and supporting utilities.
• utils/… modules are duplicated under both branches, providing camera, image, kernel‐grid, loss, and visualization utilities.
• External dependencies: PyTorch, CUDA, CMake, Colmap outputs, optional tiny-cuda-nn.
README for additional insights
• Provides a high‐level “pipeline.png” showing data feeding into GS and SDF, joint supervision, and final outputs (renderings and geometry).
• Specifies data organization under a data/ folder, and training/evaluation commands.
Elements
• Green for SDF branch components
• Gray for common utilities
• Orange for C++/CUDA submodules