Scene flow versus optical flow

WHAT THIS TESTS Whether you understand that optical flow is a projection of true motion onto the image plane, while scene flow is the underlying three-dimensional motion, and what inputs that upgrade requires.

A GOOD ANSWER COVERS Optical flow assigns each pixel a two-dimensional displacement in the image, the apparent motion in the picture. It conflates depth and motion: a near slow object and a far fast object can produce identical optical flow. Scene flow assigns each point in the scene a three-dimensional velocity vector, the dense 3D motion field of the world, separating true motion in x, y, and z. Because it lives in 3D, scene flow requires depth information that a single image cannot give. Typical inputs are a calibrated stereo pair, an RGB-D camera, or LiDAR point clouds, sometimes fused. The high-level algorithm: first recover scene geometry, depth or 3D points, at each time step; second, establish correspondences over time, often by combining optical flow in the image with stereo disparity; third, jointly optimize for the 3D motion that explains the observed image and depth changes, regularized by smoothness and piecewise-rigidity priors so nearby points move coherently. Modern methods learn this directly from stereo or point-cloud pairs.

COMMON WRONG ANSWERS Saying scene flow is just optical flow with a depth channel appended; it is a genuinely 3D motion field requiring joint reasoning. Forgetting that depth sensing is mandatory. Ignoring rigidity or smoothness priors that make the problem well posed.

LIKELY FOLLOW-UPS How does stereo disparity combine with optical flow to give scene flow? Why is the problem ill-posed without priors? How do point-cloud methods like FlowNet3D estimate it? What ambiguities remain in textureless regions?

ONE CONCRETE EXAMPLE A car approaching a stereo rig: optical flow shows the car expanding in the image, ambiguous between approach and growth. Using stereo depth at two times plus flow, scene flow recovers an explicit velocity vector pointing toward the camera in 3D, distinguishing real forward motion from mere image-plane change.