Cross-trees, Edge and Superpixel Priors-based Cost aggregation for Stereo matching

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• Nonlocal cost aggregation is crucial for stereo matching and image processing tasks.
• Existing tree-based methods like Minimum Spanning Tree (MST) and Segment-Tree (ST) have limitations in efficiency and uniqueness.
• Accurate cost aggregation is essential for reliable depth estimation and scene reconstruction.

Purpose of the Study:

• To propose a novel cross-trees structure for efficient and unique nonlocal cost aggregation.
• To introduce edge and superpixel priors to prevent false cost aggregations across depth boundaries.
• To evaluate the performance of the proposed cross-trees method against existing tree-based nonlocal algorithms.

Main Methods:

• A novel cross-trees structure, comprising a horizontal-tree and a vertical-tree, is proposed for cost aggregation.
• The cross-trees construction is independent of image properties, ensuring efficiency and uniqueness.
• Two algorithms, cross-trees with edge prior and cross-trees with superpixel prior, are developed to handle depth discontinuities.

Main Results:

• The proposed cross-trees structure provides efficient and unique tree constructions.
• The algorithms incorporating edge and superpixel priors effectively tackle false cost aggregations across depth boundaries.
• Performance evaluations on 27 Middlebury datasets demonstrate superior results compared to MST and ST methods.

Conclusions:

• The novel cross-trees structure offers a significant advancement in nonlocal cost aggregation techniques.
• The proposed priors effectively enhance the robustness of cost aggregation by respecting depth boundaries.
• The cross-trees method presents a faster and more accurate alternative for stereo matching and related computer vision applications.