Reproduce “Bilateral Grid Learning for Stereo Matching Networks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bin Xu, Yuhua Xu, Xiaoli Yang, Wei Jia, Yulan Guo
This research introduces a Cost Volume Upsampling in Bilateral Grid (CUBG) module, which enables deep stereo matching networks to achieve real-time performance while maintaining high accuracy. The CUBG module accelerates existing networks by 4-29 times and forms the core of BGNet, which processes stereo pairs at 39 frames per second with accuracy superior to prior real-time methods.
Key insights
- A differentiable bilateral grid operation can be successfully implemented within deep stereo matching networks to perform efficient, edge-preserving cost volume upsampling.
- Decoupling computationally intensive cost aggregation to a low resolution and subsequently using an edge-aware upsampling mechanism for high-resolution disparity regression is an effective strategy to balance real-time speed with high accuracy.
- Learned high-resolution image feature maps serve as more effective guidance for the bilateral grid's slicing operation compared to luma-version input images, leading to improved edge preservation and overall accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.