Automatic Depth Extraction from 2D Images Using a Cluster-Based Learning Framework.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Growing demand for 3D content contrasts with limited availability.
• Existing 2D-to-3D conversion algorithms face challenges in accuracy and automation.

Purpose of the Study:

• To develop an automatic, learning-based algorithm for 2D-to-3D image conversion.
• To leverage structural similarity between color images and their depth maps for accurate depth estimation.

Main Methods:

• A K-Nearest Neighbor framework clusters a database of color+depth images based on structural similarity.
• Prior depth maps are generated from cluster representatives and selected via feature descriptor comparison.
• Segmentation-guided filtering refines the initial depth estimation for improved accuracy.

Main Results:

• The algorithm successfully estimates depth maps for query images.
• Performance was validated on public databases against state-of-the-art methods.
• The approach demonstrates efficiency in automatic 2D-to-3D conversion.

Conclusions:

• The proposed learning-based method effectively converts 2D images to 3D by utilizing structural priors.
• This technique addresses the scarcity of 3D content by providing an automated conversion solution.
• The method shows promise for enhancing 3D media production.