Depth Transfer: Depth Extraction from Video Using Non-Parametric Sampling.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Accurate depth map generation is crucial for 3D reconstruction and visualization.
• Existing methods struggle with non-translating cameras and dynamic scenes.
• Automatic conversion of monoscopic video to stereo 3D is highly desirable.

Purpose of the Study:

• To develop a robust technique for automatically generating plausible depth maps from videos.
• To improve depth estimation in challenging scenarios where prior methods fail.
• To enable automatic conversion of monoscopic videos to stereo 3D.

Main Methods:

• Non-parametric depth sampling for depth map generation.
• Utilizing local motion cues and optical flow for video depth enhancement and temporal consistency.
• Training and evaluation using a Kinect-based system with stereoscopic videos and known depths.

Main Results:

• The proposed technique outperforms state-of-the-art methods on benchmark datasets.
• Successfully generates plausible depth maps for non-translating cameras and dynamic scenes.
• Demonstrates visually pleasing results for automatic monoscopic to stereo video conversion.

Conclusions:

• The developed technique offers a significant advancement in automatic depth map generation.
• It provides a robust solution for challenging video scenarios and 3D visualization applications.
• The method shows broad applicability, from research to feature film post-production.