How borglab/gtsfm works
High-Level Components
a) Front-End Layer
• Feature Detection/Description Module
• Multiple implementations (SIFT, SuperPoint, etc.)
• To Detector and Descriptor submodules
• Matching Module
• SuperGlue, traditional matchers
• Verification Module
• Degensac, RANSAC implementations
• Caching System
• Detector/descriptor and matcher caching
b) Middle Layer
• Data Association Module
• Connection to 3D point triangulation
• Averaging Module
• Rotation Averaging (Shonan)
• Translation Averaging (1D-SFM)
c) Back-End Layer
• Bundle Adjustment Module
• Dense Reconstruction Module (PatchMatchNet)
Data Flow Connections
• Image data flow from Loaders through the pipeline
• Feature extraction → matching → verification flow
• Point tracks formation
• Camera pose estimation flow
• Final 3D reconstruction output
Distributed Computing Layer
Show:
• Dask integration for parallel computation
• Worker distribution
• Task scheduling
Supporting Systems
Include:
• Data Loaders (various formats)
• Visualization Tools
• Evaluation Metrics
• Caching System
External Integrations
Connections to:
• GTSAM
• Nerfstudio
• COLMAP (for comparisons)
Use the following visual elements:
• Rounded rectangles for submodules
• Blue: Front-end components
• Green: Middle-layer components
• Orange: Back-end components
• Gray: Supporting infrastructure
• Purple: External integrations
Key Design Principles to Highlight:
Integration Points: Display external tool connections
• Component types
• Implementation options for each module
This design captures both the high-level architecture and the detailed module interactions while highlighting the system's key features: distributed computation, modular design, and extensive front-end capabilities.