Video-driven 2D Digital Twin

An indoor video shows the room from a camera’s perspective; a floor plan needs a shared top-down view and meaningful scale. Our team’s project connects these representations through a modular pipeline: background extraction, semantic segmentation, per-view bird’s-eye-view transformation, fusion, door and scale reasoning, and metric floor-plan generation for later movement mapping.

I owned the door-detection module and contributed candidate-mask processing, lower-endpoint estimation, and scale anchoring. The useful output is not simply “there is a door”: downstream reconstruction needs geometrically meaningful points and an anchor for scale. This connected my detection work to the coordinate and representation needs of later pipeline stages.

My technical contribution: make visual detections usable by downstream geometry.

I also contributed to per-view BEV alignment and multi-view fusion. The documented fusion stage combines registration strategies, bidirectional loop checks, robust pose-graph reasoning, and pixel/footprint consensus with uncertainty-aware outputs. These checks matter because agreement between views needs to be established, not assumed when images are merged.

Beyond individual modules, I set up and managed the shared GitHub repository and coordinated contribution structure and interfaces. The project remains in active development in the August 2026 record. My ownership is specific—repository coordination, door geometry, and fusion contributions—within a broader system built by the team.