Consistent depth maps recovery from a video sequence.

Area of Science:

• Computer Vision
• Photogrammetry
• 3D Reconstruction

Background:

• Stereo reconstruction methods struggle with image noise, occlusions, and outliers.
• Maintaining temporal coherence in dense depth maps is a significant challenge.

Purpose of the Study:

• To develop a novel method for recovering consistent depth maps from video sequences.
• To address limitations of existing multi-view stereo techniques.

Main Methods:

• A bundle optimization framework is proposed, integrating photo-consistency and statistical geometric coherence across multiple frames.
• An iterative optimization scheme initializes disparity maps with a segmentation prior, followed by bundle optimization.
• Implicit modeling of reconstruction noise and probabilistic visibility replaces explicit visibility parameter definition.

Main Results:

• The method naturally maintains temporal coherence in dense depth maps without over-smoothing.
• An efficient space-time fusion algorithm further reduces reconstruction noise post-optimization.
• The approach demonstrates robust performance on challenging video sequences.

Conclusions:

• The proposed bundle optimization framework offers a robust solution for depth map recovery from videos.
• The method effectively handles noise, occlusions, and outliers while preserving temporal consistency.
• This work advances automatic depth recovery techniques for video sequences.