Optical flow has been a cornerstone of computer vision for decades, but it only captures 2D motion in the image plane. This research introduces a method to upgrade optical flow to 3D scene flow by leveraging 'optical expansion'—the change in apparent size of objects as they move in depth. The key insight is that expansion cues, combined with traditional flow, can recover 3D motion vectors without requiring stereo or depth sensors. The method is evaluated on standard benchmarks and shows significant improvements in 3D motion estimation accuracy. For computer vision researchers and engineers, this work opens new possibilities for monocular 3D scene understanding, with direct applications in autonomous navigation, robotic manipulation, and AR/VR. The approach is computationally efficient, making it suitable for real-time systems. This signal is particularly relevant as the industry pushes toward more capable vision systems that can operate in unconstrained environments.
A novel method that uses optical expansion to extend 2D optical flow to 3D scene flow, enabling depth-aware motion estimation.