CamGraph
CamGraph: Level-of-Detail Visualization of Camera View Graphs for Novel View Synthesis. A tool for inspecting the camera viewpoints behind large-scale 3D reconstructions without the visual clutter of showing every camera at once.
Vincent Limardi, Ming-Hsien Huang, Kai-Wen Hsiao, Shih-Hsuan Hung
Overview
CamGraph helps researchers make sense of the hundreds or thousands of camera viewpoints used to reconstruct a 3D scene with modern Novel View Synthesis methods like 3D Gaussian Splatting. From Structure-from-Motion outputs, it builds a weighted camera view graph encoding co-visibility and pose geometry, hierarchically clusters cameras into a tree of groups via spectral clustering, and uses a viewpoint-aware level-of-detail viewer to surface the camera groups and input views most relevant to wherever you're looking — instead of cluttering the scene with every camera frustum at once.
Technologies
Results
Comparison between a conventional frustum viewer (a, cluttered) and CamGraph (b). The case study (c-e) traces a degraded, blurred 3D Gaussian Splatting reconstruction back to the input camera views responsible for it.
The Challenges
- Dense captures with hundreds to thousands of cameras overwhelm conventional frustum-based viewers with visual clutter
- Existing tools (COLMAP, Nerfstudio Viewer) show cameras as isolated frustums/icons with no sense of how views relate to each other
The Solutions
- Built a weighted camera view graph from SfM outputs encoding co-visibility and pose geometry
- Applied recursive spectral clustering to organize cameras into a tree-structured hierarchy
- Designed a level-of-detail viewer that reveals coarse camera groups when zoomed out and fine camera detail on local inspection
Key Outcomes
- Accepted as a poster at SIGGRAPH Asia 2026 (Kuala Lumpur, Malaysia)
- Validated on a 1,000+ camera outdoor case study and a 12-participant user study with a 100% task success rate