Open Arsenal

The geometry was always there.

HelixPlan is an open-source spatial intelligence system designed to reconstruct accurate architectural floorplans from simple walkthrough video and minimal user-provided dimensions. It bridges computer vision, geometry processing, and machine learning to transform real-world environments into structured digital representations. Instead of relying on manual drafting or expensive scanning equipment, HelixPlan infers spatial structure directly from motion and visual observation.

At its core, HelixPlan processes video through a multi-stage pipeline that includes SLAM-based reconstruction, depth estimation, and semantic scene understanding. It builds a 3D representation of the environment and then converts that structure into a clean, navigable 2D floorplan. The system identifies key architectural elements such as walls, doors, windows, closets, and stairs, and organizes them into a coherent spatial graph.

A key feature of HelixPlan is its Structural Constraint Engine, which ensures that generated layouts remain physically and architecturally valid. Doors must connect meaningful spaces, rooms must be properly enclosed, and staircases must align with realistic vertical transitions. This prevents the system from producing geometrically plausible but physically impossible structures.

To improve stability and accuracy, HelixPlan includes a Temporal Consistency Model that reduces drift and inconsistency across video frames. It stabilizes geometry over time, ensuring that detected structures remain consistent throughout the reconstruction process. Additionally, a Measurement Verification Layer cross-checks inferred dimensions against architectural standards and user-provided scale inputs, improving reliability.

Finally, HelixPlan continuously improves through a Dataset Generation Pipeline that converts real-world walkthroughs into structured training data. Each processed environment can contribute back into the system as labeled geometry, enabling iterative refinement of its models. The result is a continuously evolving open system for understanding and reconstructing built space.