VGGT-SLAM

Why we included this project

VGGT-SLAM turns the feed-forward VGGT scene reconstruction model into a full dense SLAM pipeline that works from ordinary monocular video without camera calibration. It breaks long sequences into submaps and aligns them on the SL(4) manifold, which handles the projective ambiguity that plain similarity transforms miss. That approach lets it reconstruct large, room-scale scenes from hundreds or thousands of RGB frames that a single VGGT pass couldn't fit in GPU memory, with no extra training and no known intrinsics. The repo includes a script that produces an incremental map visualization, and optional open-set 3D object detection via SAM 3 and Perception Encoder is there for vision researchers and robotics engineers prototyping dense mapping. Setup automates downloading the third-party model weights, so you can get a reconstruction running fairly quickly if you have the GPU headroom.