Reproduce “FractalMamba++: Scaling Vision Mamba Across Resolutions via Hilbert Fractal Geometry” 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
Bo Li, Haoke Xiao, Lv Tang
Researchers from vivo Mobile Communication Co., Ltd., Kuaishou Technology, and the University of Alberta developed FractalMamba++, a Vision Mamba backbone that uses Hilbert fractal geometry to achieve robust resolution scalability. The model consistently outperforms existing Vision Mamba variants on high-resolution image classification, object detection, semantic segmentation, and change detection tasks by enhancing spatial continuity, long-range information propagation, and positional encoding.
Key insights
- A unified geometric principle, such as the self-similar structure of the Hilbert curve, can deterministically guide the design of multiple backbone components for consistent and resolution-invariant performance.
- Preserving 2D spatial locality during 2D-to-1D sequence conversion is more effective for high-resolution generalization in Vision Mamba models than relying solely on multi-scale training.
- Deterministic, geometry-derived skip connections can efficiently counteract information fading in recurrent SSMs, allowing for better global context modeling without increasing asymptotic sequence complexity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.