Active Disparity Sampling for Stereo Matching With Adjoint Network.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Sparse depth measurements are used to guide dense disparity estimation in stereo vision.
• Existing methods assume random sampling, leading to inefficient use of data in challenging and well-estimated regions.

Purpose of the Study:

• To introduce an Active Disparity Sampling (ADS) problem for optimizing depth measurement utility.
• To enhance dense disparity estimation performance with arbitrary sampling budgets.

Main Methods:

• Learning an Adjoint Network to predict pixel-wise disparity quality for deep stereo models.
• Implementing a hard-soft prior supervision for hierarchical quality map learning.
• Proposing a Bayesian optimized sampling policy guided by disparity quality.

Main Results:

• Demonstrated effectiveness across various stereo architectures and datasets.
• Outperformed existing fixed and adaptive sampling methods at different sampling rates.
• Showcased substantial improvements when generalized to heterogeneous unseen domains.

Conclusions:

• Active Disparity Sampling enhances the utility of sparse depth measurements for dense disparity estimation.
• The proposed method offers a flexible and effective approach to sampling strategies in stereo vision.
• The approach generalizes well to diverse and previously unseen data domains.