Minimum Spanning Forest with Embedded Edge Inconsistency Measurement Model for Guided Depth Map Enhancement.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Guided depth map enhancement often relies on Markov Random Field (MRF) models.
• Existing MRF methods assume color-depth edge consistency, leading to artifacts like texture-copying and blurred edges.
• Previous approaches often ignore local structural information in depth maps.

Purpose of the Study:

• To propose a novel MRF-based method for guided depth map enhancement.
• To address limitations of existing methods, specifically texture-copying artifacts and blurred depth edges.
• To improve depth map super-resolution and completion by preserving depth edges more effectively.

Main Methods:

• A new MRF-based approach is presented that computes affinities using pixel distances within a Minimum Spanning Tree (Forest) space.
• Edge weights within each Minimum Spanning Tree are calculated using an explicit edge inconsistency measurement model.
• A bandwidth adaptation scheme is incorporated to enhance noise tolerance and depth edge preservation.

Main Results:

• The proposed method significantly mitigates texture-copying artifacts.
• Depth edges are better preserved compared to existing techniques.
• Evaluations on synthetic and real datasets (Middlebury, ToF-Mark, NYU) demonstrate improved performance in depth map super-resolution and completion.

Conclusions:

• The novel MRF method effectively enhances depth maps by preserving edges and reducing artifacts.
• The use of Minimum Spanning Trees and an edge inconsistency model offers a significant advancement.
• The method shows superior qualitative and quantitative results against 16 state-of-the-art techniques.