One-eyed stereo: a general approach to modeling 3-d scene geometry

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Single 2D images are ambiguous representations of 3D scenes.
• Human vision excels at inferring depth from 2D images.
• Existing "shape from..." methods (e.g., shading, texture, contour) use distinct assumptions.

Purpose of the Study:

• To unify "shape from..." methods under a single framework.
• To demonstrate that these methods can be translated into conventional stereo formalism.
• To present a new algorithm for depth recovery from single images.

Main Methods:

• Reinterpreting "shape from..." assumptions as a virtual image.
• Structuring depth recovery as a stereo problem using a perspective and an hypothesized orthographic image.
• Developing a novel stereo reconstruction algorithm.

Main Results:

• Demonstrated that "shape from..." methods are equivalent to stereo vision.
• Showcased the possibility of using a perspective and a virtual orthographic image for depth recovery.
• Presented a new algorithm capable of this stereo reconstruction task.

Conclusions:

• "Shape from..." methods can be unified within a stereo vision framework.
• Depth recovery from single images is achievable by creating a stereo pair with a virtual image.
• The proposed algorithm offers a new method for single-image 3D reconstruction.