Digging Into Uncertainty-Based Pseudo-Label for Robust Stereo Matching

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Stereo matching methods struggle with domain shift and imbalanced disparity, limiting dataset generalization.
• Adapting to new domains typically requires costly ground-truth data, which is often impractical.

Purpose of the Study:

• To develop a robust stereo matching approach that addresses domain shift and limited ground truth data.
• To leverage uncertainty estimation for improved disparity distribution and model adaptation.

Main Methods:

• Employed pixel-level uncertainty estimation to dynamically adjust disparity search spaces, pruning unlikely correspondences.
• Introduced uncertainty-based pseudo-labeling (pixel-level and area-level) to adapt pre-trained models to new domains using sparse, reliable labels.
• Validated the method's effectiveness in cross-domain, adaptation, and joint generalization scenarios.

Main Results:

• Achieved 1st place in the stereo task at the Robust Vision Challenge 2020.
• Demonstrated strong cross-domain, adaptation, and joint generalization capabilities.
• Extended uncertainty-based pseudo-labels to unsupervised monocular depth estimation, achieving performance comparable to supervised methods.

Conclusions:

• Uncertainty estimation offers a robust solution for stereo matching challenges, particularly domain shift and data scarcity.
• The proposed uncertainty-based pseudo-labeling technique effectively bridges domain gaps and enables unsupervised learning for depth estimation.