Reproduce “IGEV++: Iterative Multi-range Geometry Encoding Volumes for Stereo Matching” 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
Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Junda Cheng, Chunyuan Liao, Xin Yang
IGEV++ presents an iterative deep network for stereo matching that integrates cost volume filtering with iterative optimization through Multi-range Geometry Encoding Volumes (MGEV). The architecture achieves state-of-the-art accuracy on Scene Flow, Middlebury, KITTI, and ETH3D benchmarks, reducing the 2-pixel outlier rate on Middlebury by 31.9% compared to RAFT-Stereo while demonstrating robust performance in ill-posed regions and for large disparities.
Key insights
- Cost volume filtering and iterative optimization methods possess complementary advantages, where filtering provides robust contextual information and iterative methods offer efficiency for large disparities.
- Encoding geometry at multiple ranges and granularities (fine-grained for details, coarse-grained for large disparities/ill-posed regions) is more effective than single-granularity approaches.
- Providing comprehensive, aggregated geometry features, rather than just local correlations, to an iterative refinement module can drastically improve convergence speed and accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.