LingBot-Map
Feed-forward transformer that rebuilds 3D scenes from streaming video
What it is
LingBot-Map is a research model and codebase for streaming 3D reconstruction that estimates camera poses and dense geometry frame by frame from video, relying on anchor context, a window of reference poses, and a memory of past trajectory to keep drift in check. Paged KV cache attention keeps inference around 20 FPS over sequences of more than 10,000 frames. The repo ships weights on Hugging Face and ModelScope, an interactive browser viewer, an offline rendering pipeline, and benchmark scripts for datasets such as KITTI and Oxford Spires. Computer vision and robotics researchers can use it to map long indoor, outdoor, or aerial walkthroughs.