Reproduce “Learning Intra-view and Cross-view Geometric Knowledge 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
Rui Gong, Weide Liu, Zaiwang Gu, Xulei Yang, Jun Cheng
ICGNet enhances deep stereo matching by integrating intra-view and cross-view geometric knowledge, derived from pre-trained local feature matching models, during training. The framework improves state-of-the-art networks like IGEV-Stereo, achieving 0.447 px EPE on SceneFlow and 1.41% D1-all on KITTI 2015, while also boosting cross-domain generalization without inference-time overhead.
Key insights
- Both intra-view and cross-view geometric knowledge are crucial for robust stereo matching; intra-view knowledge particularly benefits occluded regions, while cross-view knowledge enhances non-occluded areas.
- Powerful, pre-trained local feature matching pipelines (like SuperPoint and SuperGlue) can effectively serve as 'teachers' to transfer rich geometric understanding to stereo matching networks during training.
- Explicitly learning geometric knowledge significantly improves cross-domain generalization, helping stereo models trained on synthetic data perform better on real-world datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.