A Wavelet-based Stereo Matching Framework for Solving Frequency Convergence Inconsistency

Xiaobao Wei, Jiawei Liu, Dongbo Yang, Junda Cheng, Changyong Shu, Wei Wang

Keywords: real time stereo matching, depth estimation, disparity

Abstract:

Through an in-depth analysis the underlying cause of the limited performance in iterative stereo matching methods: \textbf{frequency convergence inconsistency}, we propose a novel plug-and-play module named Wavelet-Stereo for this inherent flaw. Specifically, we first summarize the convergence characteristics of distinct frequency components and designed a specialized dual-branch architecture.
The high-frequency branch rapidly captures detailed context by a unet, while the low-frequency branch progressively refines the textureless regions throughout the iteration.
These two branches interact via a carefully designed high-frequency preservation update operator and predict the disparity, achieving synchronous optimization of both high and low frequency regions.
Extensive experiments demonstrate that our Wavelet-Stereo outperforms the state-of-the-art methods and ranks $1^{st}$ on SceneFlow, ETH3D, KITTI 2015 and KITTI 2012 online leaderboards for almost all metrics. Our work not only uncovers the phenomenon of frequency convergence inconsistency for the first time, but also provides an effective solution and paves the way for new research directions in stereo matching.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 12103