How compvis/edgs works
Project Type and Purpose
• EDGS is an open-source research codebase for efficient 3D scene reconstruction using Gaussian Splatting. It provides both a CLI/Notebook training pipeline and a Gradio‐based interactive demo.
File Structure Analysis
• Top-level scripts:
• train.py, full_eval.py – training and evaluation entrypoints
• gradio_demo.py – launches interactive web demo
• install.sh – environment setup
• Directories:
• configs/ – Hydra YAML configurations
• source/ – core implementation (initialization, training, networks, losses, utils)
• submodules/gaussian-splatting & submodules/RoMa – third-party libs for rasterization and correspondence models
• notebooks/ – example Jupyter workflow
• assets/ – media examples and GIFs
• Key patterns:
• Hydra config management
• Modular pipeline: preprocessing → initialization → training → evaluation → visualization
• External dependencies: pycolmap, PyTorch, WandB, Gradio
High-Level Components
a. Data Layer (Input/Output)
• Input sources:
• Local dataset folders (images + COLMAP sparse)
• Single video file ingestion
• Outputs:
• Trained Gaussian model files
• Rendered videos/previews
• Logs/metrics
b. Preprocessing Module
• COLMAP / pycolmap wrapper (source/utils_preprocess.py)
• Image/video → camera poses + sparse point cloud
c. Initialization Module
• corr_init (source/corr_init.py)
• Uses RoMa submodule to extract 2D correspondences and initialize Gaussians
d. Training Engine
• trainer (source/trainer.py) orchestrates epochs
• networks (source/networks.py) defines UDFs for Gaussian Splatting
• losses (source/losses.py), data_utils, timer
• Hydra config injection (train.gs_epochs, train.no_densify, etc.)
e. Evaluation Module
• full_eval.py calls metrics.py across datasets
f. Visualization & Demo
• source/visualization.py for plots and Open3D previews
• gradio_demo.py provides frontend (Gradio)
g. Configuration Manager
• Hydra (configs/base.yaml, train.yaml) drives parameterization
h. External Services & Libraries
• WandB for logging (optional)
• Hugging Face Spaces for hosted demo
• GPU/CUDA runtime environment
Relationships and Data Flow
User → (CLI / Notebook / Gradio) → Preprocessing → Initialization → Training → Evaluation → Visualization → User
• Config Manager ties into each module to supply parameters
• Logging service (WandB) receives data from Training and Evaluation
• Submodules are dependencies of Initialization and Training blocks
Architectural Patterns & Design Principles
• Modular pipeline stages (separation of concerns)
• Config-driven execution (Hydra) for reproducibility
• Plugin architecture via submodules (gaussian-splatting, RoMa)
• Multi-interface support (CLI, Notebook, Web UI)