π^4: A Foundation Model for 4D Scene Geometry and Motion
π⁴ reconstructs scene geometry and tracks motion in one pass with roughly 500 million parameters.
It accepts a single image, unordered photos or video without requiring camera calibration, recovering dense geometry, camera poses and metric scale. For multiple images, it also produces 2D and 3D point tracks and moving-object masks. The authors report the best overall average rank on geometry benchmarks, alongside competitive tracking and segmentation results. They also demonstrate scenes with thousands of input images, despite training on at most 64 views at a time.
HF Daily Papers' note
It accepts a single image, unordered photos or video without requiring camera calibration, recovering dense geometry, camera poses and metric scale. For multiple images, it also produces 2D and 3D point tracks and moving-object masks. The authors report the best overall average rank on geometry benchmarks, alongside competitive tracking and segmentation results. They also demonstrate scenes with thousands of input images, despite training on at most 64 views at a time.
HF Daily Papers' note
score 5