Reproduce “Unblur-SLAM: Dense Neural SLAM for Blurry Inputs” 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
Qi Zhang, Denis Rozumny, Francesco Girlanda, Sezer Karaoglu, Marc Pollefeys, Theo Gevers, Martin R. Oswald
Unblur-SLAM introduces an online SLAM pipeline that robustly reconstructs sharp 3D scenes and estimates camera poses from inputs affected by both motion and defocus blur. It achieves this by adaptively processing frames based on blur levels and integrating a physics-constrained deblurring network, leading to superior reconstruction quality (e.g., PSNR 29.49 dB on Deblur-NeRF) and more accurate pose estimation (e.g., ATE 0.336m on TUM) compared to existing methods.
Key insights
- Adaptive blur detection and selective processing significantly reduce computational overhead while maintaining accuracy, making the system more efficient for mixed-quality inputs.
- Training a deblurring network with physics-constrained blur formation models and a two-stage approach (mid-exposure frame extraction, then defocus robustness) is crucial for 3D-consistent deblurring.
- Effectively combining a preliminary feed-forward deblurring model with 3D Gaussian-based multi-view blur kernel learning allows for robust handling and reconstruction from diverse blur types.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.