3/4 MVTracker assumes calibrated cameras with depth (e.g., from Kinect); when not available, cameras and depth can be estimated with off-the-shelf models such as DUSt3R or VGGT.
2/4 Key idea: fuse multi-view features into a unified 3D point cloud, link them with kNN-based correlations, and refine trajectories iteratively with transformer attention.
1/4 🚀 We’re excited to release MVTracker (ICCV 2025 Oral), the first data-driven multi-view 3D point tracker. MVTracker tracks arbitrary 3D points across multiple cameras, handling occlusions and varied camera setups without per-sequence optimization.
