Reproduce “UniStitch: Unifying Semantic and Geometric Features for Image Stitching” 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
Yuan Mei, Lang Nie, Kang Liao, Yunqiu Xu, Chunyu Lin, Bin Xiao
UniStitch introduces a unified framework for image stitching that integrates both geometric and semantic features, achieving superior alignment accuracy and visual quality across diverse in-domain and out-of-distribution scenes. The approach significantly improves performance, reaching 25.07 mPSNR and 0.813 mSSIM on the UDIS-D dataset and showing enhanced robustness in unseen environments, while also boosting efficiency for high-resolution images.
Key insights
- The synergistic integration of discrete geometric features and continuous semantic features provides superior robustness and generalization across diverse stitching scenarios.
- Adaptive weighting of modalities through an Adaptive Mixture of Experts and a Modality Robustifier allows the system to intelligently leverage the most reliable features in varying conditions.
- Reformulating high-resolution Thin-Plate Spline (TPS) transformations with Free-Form Deformation (FFD) significantly improves memory efficiency and inference speed while maintaining alignment precision.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.