Advanced Vision and Imaging

Computers learn to perceive the three-dimensional structure of a scene from flat, two-dimensional images by analyzing cues like stereo disparity between camera pairs, motion across video frames, and subtle shading patterns in a single photograph. These techniques underpin autonomous vehicles, robotic manipulation, augmented reality, and medical imaging, wherever a machine needs to understand where objects are in space, not just what they look like. Much of the current research centers on training deep convolutional networks to estimate depth and motion without requiring expensive labeled ground-truth data, a direction known as unsupervised or self-supervised learning. Key open questions include how to make these models generalize reliably across unfamiliar environments and lighting conditions, and how to fuse information from multiple views or sensor modalities to achieve the accuracy and robustness that safety-critical applications demand.

Works

107,202

Total citations

1,822,397

Keywords

Stereo VisionDepth EstimationOptical FlowConvolutional NetworksMulti-View StereoUnsupervised Learning