Scene flow is the dense or semi-dense three-dimensional (3D) motion field of a physical environment captured over time. While 2D optical flow only tracks pixel movements across an image plane, scene flow computes actual 3D displacement vectors for every point or pixel in a scene. It functions as a core low-level perception task in computer vision and robotics. Core Concept: Optical Flow vs. Scene Flow Optical Flow: Tracks 2D displacement
of pixels across successive video frames. It lacks depth data and struggles with scale or camera rotation. Scene Flow: Estimates 3D displacement
for real-world points between consecutive time steps. It provides a complete understanding of structural and geometric change. How Scene Flow is Estimated
The input data dictates the mathematical approach used to solve for scene flow: 1. Image-Based (Stereo/RGB-D Cameras)
This method fuses 2D optical flow with temporal disparity changes. By aligning sequential frames from left and right cameras, models solve for 3D structure and 3D motion fields simultaneously. Object Scene Flow for Autonomous Vehicles – CVF Open Access
Leave a Reply