Reproduce “DeflareMamba: Hierarchical Vision Mamba for Contextually Consistent Lens Flare Removal” 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
Yihang Huang, Yuanfei Huang, Junhui Lin, Hua Huang
Researchers from Beijing Normal University developed DeflareMamba, a hierarchical Vision Mamba model for lens flare removal that achieves a PSNR of 27.78 dB and SSIM of 0.899 on the Flare7K++ dataset. The model also improves performance in downstream object detection and vision-language alignment tasks by providing contextually consistent flare-free images.
Key insights
- U-shaped architectures are effective for Mamba-based vision models to address the long-term decay property and process high-resolution images efficiently.
- Integrating window-based local processing with multi-directional selective scanning significantly enhances Mamba's ability to retain fine-grained local details and contextual consistency in 2D images.
- Generating multi-scale sub-sequences through stride sampling provides an adaptive and efficient method for Mamba to model long-range dependencies across an image without losing resolution.
- The combination of local-enhanced and hierarchical scanning within a U-shaped Mamba framework allows for comprehensive contextual consistency, crucial for removing complex lens flare artifacts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.