Reproduce “Diffuse the object, keep its label: curating detector training data from a few unlabeled photographs via VLM-built 3D vegetation scenes” 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
Mario Malizia, Marnix Enting, Rob Haelterman, Ken Hasselmann
Researchers from the Royal Military Academy, Brussels, developed a five-stage pipeline for generating labeled object detection training data from minimal unlabeled photographs of a target site, specifically for camouflaged objects in vegetation. This method produced synthetic datasets that, when used to train a YOLOv11-L detector, yielded a macro-F1 score of 0.584 "," 0.025 in cross-site generalization, exceeding the 0.393 "," 0.021 achieved by training on a larger real-world dataset from a different site.
Key insights
- Unsupervised site adaptation with generated data can yield better cross-site generalization than training on real, labeled data from a different source site.
- A "graded mask-lock" mechanism is essential for effective appearance transfer to rare objects; it allows photorealistic texturing while preserving object geometry and labels, significantly improving detection over hard locks or unrestricted diffusion.
- Standard in-domain synthetic validation metrics or quality-metric probes are unreliable indicators of real-world cross-site generalization, emphasizing the need for direct evaluation on target domain data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.