How gmum/medgs works
MedGS trains a folded-Gaussian representation of medical image frames, optionally learning image and segmentation heads together, then renders interpolated views or extracts a mesh. The main workflow is dataset ingestion → scene/camera setup → Gaussian optimization → rendering and mesh output. The README documents the user-facing modes; the source index and excerpts support the core model, scene, renderer, and utility connections. Some rendering and mesh internals were not sampled, so those edges rely on README statements rather than claims about their internal implementation.
Main entry points: gaussian_model.py Folded Gaussians · train.py Training modes · render.py View rendering · __init__.py Gaussian renderer
Components 22
Training workflow
- Training modes
train.py - Training configuration
arguments/__init__.py
Scene and representation
- Dataset readers
models/scenes/dataset_readers.py - Scene setup
scene/__init__.py - Camera transforms
scene/cameras.py - Scene data readers
scene/dataset_readers.py - COLMAP loading
scene/colmap_loader.py - Folded Gaussians
scene/gaussian_model.py - Point Gaussian model
models/flat_splatting/scene/points_gaussian_model.py
Rendering and outputs
- Gaussian renderer
gaussian_renderer/__init__.py - Background renderer
gaussian_renderer/bg_fg_renderer/__init__.py - View rendering
render.py - Mesh extraction
utils/mesh_utils.py - Mesh evaluation
evaluate.py - Rendered views
- Mesh results
Numerical support
- Training losses
utils/loss_utils.py - Geometry utilities
utils/graphics_utils.py - Tensor utilities
utils/general_utils.py - Spherical harmonics
utils/sh_utils.py
Other components
- Researcher
- Medical frames
Connections 25
- Researcher → Training modes: starts training
README.md - Medical frames → Dataset readers: provides frames
README.md - Training modes → Dataset readers: selects data
README.md - Dataset readers → Scene data readers: uses base readers
models/scenes/dataset_readers.py - Scene data readers → COLMAP loading: loads COLMAP
scene/dataset_readers.py - Dataset readers → Geometry utilities: uses camera geometry
models/scenes/dataset_readers.py - Training modes → Scene setup: creates scene
train.py - Scene setup → Folded Gaussians: initializes model
scene/__init__.py - Scene setup → Camera transforms: builds cameras
scene/__init__.py - Training modes → Folded Gaussians: optimizes geometry
train.py - Training modes → Training losses: uses losses
train.py - Training modes → Training configuration: reads options
README.md - Point Gaussian model → Folded Gaussians: extends model
models/flat_splatting/scene/points_gaussian_model.py - Gaussian renderer → Folded Gaussians: renders model
gaussian_renderer/__init__.py - Gaussian renderer → Spherical harmonics: evaluates colors
gaussian_renderer/__init__.py - Background renderer → Folded Gaussians: renders model
gaussian_renderer/bg_fg_renderer/__init__.py - View rendering → Gaussian renderer: renders views
render.py - Researcher → View rendering: requests rendering
README.md - View rendering → Rendered views: writes views
README.md - View rendering → Folded Gaussians: loads checkpoint
render.py - Mesh extraction → Mesh results: produces mesh
README.md - Mesh evaluation → Mesh extraction: uses mesh utilities
evaluate.py - Folded Gaussians → Tensor utilities: uses tensor helpers
scene/gaussian_model.py - Scene data readers → Geometry utilities: uses transforms
scene/dataset_readers.py - Gaussian renderer → Training losses: uses loss utilities
gaussian_renderer/__init__.py