Two cloud-based cues for estimating scene structure and camera calibration.

Area of Science:

• Computer Vision
• Photogrammetry
• Computational Imaging

Background:

• 3D scene geometry estimation is crucial for various applications.
• Traditional methods often rely on active sensors or controlled lighting conditions.
• Utilizing natural phenomena like cloud shadows offers a passive and potentially cost-effective approach.

Purpose of the Study:

• To develop and evaluate algorithms for 3D scene geometry estimation using cloud shadows.
• To explore the use of stochastically structured light from cloud shadows for geometric reconstruction.
• To estimate scene structure and focal length from static camera video data.

Main Methods:

• Algorithms utilize time series of pixel intensity values from video captured by a static outdoor camera.
• Two primary cues are employed: spatial proximity of pixels and motion of cloud shadows.
• The spatial cue leverages the likelihood of nearby points being simultaneously under cloud cover.
• The cloud motion cue provides linear constraints on scene structure, with inherent ambiguities addressed by combining cues.

Main Results:

• The proposed methods successfully estimate scene structure and focal length.
• Combining spatial and motion cues overcomes ambiguities in cloud shadow-based reconstruction.
• Evaluation on real outdoor scenes demonstrates the efficacy of the algorithms.

Conclusions:

• Cloud shadows can be effectively utilized as a source of stochastically structured light for 3D scene geometry.
• The developed algorithms offer a novel approach to passive 3D reconstruction.
• This technique shows promise for applications requiring outdoor scene understanding without specialized equipment.