Advanced Vision and Imaging

Machines that can infer the three-dimensional structure of a scene from flat images must solve problems that human visual systems handle effortlessly but that remain genuinely hard to formalize: how far away is an object, how is it moving, and how do slightly different viewpoints encode depth? Researchers approach these questions through methods ranging from stereo vision—comparing two offset camera views much as our eyes do—to monocular depth estimation, where a single image must yield geometric information with no direct stereo signal to draw on. Convolutional networks have substantially advanced what is achievable, particularly when paired with unsupervised or self-supervised training strategies that sidestep the need for dense ground-truth depth labels, which are costly to collect at scale. Active open questions include how to make these systems generalize reliably across environments they were not trained on, and how to tightly integrate visual geometry with real-time robot control tasks like visual servoing.

Works

108,018

Total citations

1,831,885

Keywords

Stereo VisionDepth EstimationOptical FlowConvolutional NetworksMulti-View StereoUnsupervised Learning