Reproduce “Probabilistic Modeling of Disparity Uncertainty for Robust and Efficient Stereo Matching” on OpenResearch

About this paper

Wenxiao Cai, Dongting Hu, Ruoyan Yin, Jiankang Deng, Huan Fu, Wankou Yang, Mingming Gong

This paper presents an uncertainty-aware stereo matching framework that efficiently estimates and disentangles data (aleatoric) and model (epistemic) uncertainty. The approach utilizes ordinal regression for data uncertainty and kernel regression for efficient model uncertainty estimation, demonstrating improved uncertainty quantification and disparity prediction accuracy across multiple benchmarks.

Key insights

• Defining uncertainty via Bayes risk allows for the disentanglement of total uncertainty into inherent data uncertainty (aleatoric) and model knowledge uncertainty (epistemic).
• Modeling the full disparity probability distribution using ordinal regression enables robust estimation of data uncertainty (variance of the PMF).
• Model uncertainty can be efficiently estimated using Kernel Regression on feature embeddings, avoiding the high computational cost of traditional methods like deep ensembles or MC dropout.

Summary by alphaXiv.