Reproduce “SegFly: A Dataset and 2D-3D-2D Paradigm for Aerial RGB-Thermal Semantic Segmentation at Scale” 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
Markus Gross, Sai Bharadhwaj Matha, Rui Song, Viswanathan Muthuveerappan, Conrad Christoph, Julius Huber, Daniel Cremers
A new geometry-driven 2D-3D-2D paradigm enables the efficient creation of SegFly, a large-scale dataset for aerial RGB-Thermal semantic segmentation. This framework automatically generates dense pseudo ground-truth labels for over 20,000 RGB images and 15,000 RGB-T pairs from a minimal set of manual annotations, achieving 91% pixel accuracy for RGB and 88% for thermal pseudo-labels.
Key insights
- Multi-view geometric redundancy in aerial imagery can amortize manual annotation costs across vast numbers of images, enabling large-scale dataset generation.
- 3D reconstruction provides a robust intermediate domain for both generating dense pseudo-labels and accurately registering multi-modal imagery (RGB-Thermal) without hardware synchronization.
- High-quality, geometry-derived pseudo-supervision from large-scale datasets substantially improves the performance and generalization capabilities of both conventional and vision foundation semantic segmentation models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.