How zju3dv/pgsr works
PGSR reconstructs surfaces from multi-view RGB imagery using planar Gaussian splatting. The primary workflow prepares a dataset, loads camera and scene data, trains an adaptive Gaussian representation, then renders views and extracts a mesh; DTU and Tanks and Temples evaluation are additional product branches. The source index supports the principal training, scene, renderer, and rasterization links. Several dataset-specific scripts are not sampled, so their internal calls are not asserted.
Main entry points: train.py Training loop · gaussian_renderer/__init__.py Plane renderer · scene/__init__.py Scene and cameras · gaussian_model.py Gaussian model
Components 24
Dataset preparation
- Dataset conversion
scripts/preprocess/convert.py - TnT conversion
scripts/preprocess/convert_tnt.py
Reconstruction runtime
- Scene readers
scene/dataset_readers.py - COLMAP loader
scene/colmap_loader.py - Scene and cameras
scene/__init__.py - Camera utilities
utils/camera_utils.py - Camera model
scene/cameras.py - Training loop
train.py - Gaussian model
scene/gaussian_model.py - Training losses
utils/loss_utils.py - Geometry utilities
utils/graphics_utils.py - Optimization utilities
utils/general_utils.py - Spherical harmonics
utils/sh_utils.py
Rendering and export
- Plane renderer
gaussian_renderer/__init__.py - Differentiable rasterizer
submodules/diff-plane-rasterization/diff_plane_rasterization/__init__.py - Rasterization API
submodules/diff-plane-rasterization/rasterize_points.h - Render and mesh export
render.py - TnT rendering
scripts/render_tnt.py
Evaluation
- Image metrics
metrics.py - DTU evaluation
scripts/eval_dtu/eval.py - TnT evaluation
scripts/tnt_eval/evaluation.py - LPIPS metric
lpipsPyTorch/modules/networks.py
Other components
- Researcher
- Multi-view images
Connections 30
- Researcher → Dataset conversion: prepares data
README.md - Researcher → TnT conversion: prepares TnT
README.md - Multi-view images → Dataset conversion: converts images
README.md - Dataset conversion → Scene readers: produces scene data
- TnT conversion → Scene readers: produces scene data
- Scene readers → COLMAP loader: loads reconstruction
scene/dataset_readers.py - Scene and cameras → Scene readers: loads scene info
scene/__init__.py - Scene and cameras → Camera utilities: builds cameras
scene/__init__.py - Camera utilities → Camera model: constructs
- Researcher → Training loop: starts training
README.md - Training loop → Scene and cameras: creates scene
train.py - Training loop → Gaussian model: optimizes representation
train.py - Training loop → Plane renderer: renders training views
train.py - Training loop → Training losses: computes objectives
train.py - Training loop → Geometry utilities: warps patches
train.py - Gaussian model → Optimization utilities: uses transforms
scene/gaussian_model.py - Gaussian model → Spherical harmonics: uses harmonics
scene/gaussian_model.py - Plane renderer → Gaussian model: reads Gaussians
gaussian_renderer/__init__.py - Plane renderer → Camera model: reads camera
- Plane renderer → Differentiable rasterizer: rasterizes
gaussian_renderer/__init__.py - Differentiable rasterizer → Rasterization API: calls CUDA API
submodules/diff-plane-rasterization/diff_plane_rasterization/__init__.py - Researcher → Render and mesh export: renders and exports
README.md - Render and mesh export → Scene and cameras: loads saved scene
render.py - Render and mesh export → Plane renderer: renders views
render.py - Researcher → TnT rendering: renders TnT
README.md - TnT rendering → Scene and cameras: loads scene
scripts/render_tnt.py - TnT rendering → Plane renderer: renders views
scripts/render_tnt.py - Researcher → DTU evaluation: evaluates reconstruction
README.md - Researcher → TnT evaluation: evaluates reconstruction
README.md - Image metrics → LPIPS metric: uses features
metrics.py