Why Stereo Vision is Not Always About 3D Reconstruction
W. Ericl . Grimson · DSpace@MIT (Massachusetts Institute of Technology) · 1993
A common assumption of stereo vision researchers is that the goal of stereo is to compute explicit 3D information about a scene, to support activities suchasnavigation, hand-eye coordination and object recognition. This paper suggests reconsidering what is required of a stereo algorithm, in light of the needs of the task that uses its output. We show that very accurate camera calibration is needed to reconstruct accurate 3D distances, and argue that often it may be difficult to attain and maintain such accuracy.We further argue that for tasks such as object recognition, separating object from background is of central importance. We suggest that stereo can help with this task, without explicitly computing 3D information. We provide a demonstration of a stereo algorithm that supports separating figure from ground through attentive fixation on key features.