Reproduce “Leveraging Synthetic Data for Enhancing Egocentric Hand-Object Interaction Detection” 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
Rosario Leonardi, Antonino Furnari, Francesco Ragusa, Giovanni Maria Farinella
Researchers at the University of Catania developed a pipeline to generate synthetic data for egocentric hand-object interaction (HOI) detection, creating the HOI-Synth benchmark. Their work shows that semi-supervised domain adaptation using this synthetic data can achieve comparable or superior performance to 100% real-data training with only 10-25% labeled real data, significantly reducing annotation needs.
Key insights
- Despite a significant initial domain gap, effective domain adaptation strategies (like UDA) can dramatically improve model performance when trained on synthetic data, boosting Overall AP by +23.45% on VISOR.
- Semi-supervised domain adaptation (SSDA) with synthetic data can match or exceed the performance of models trained on 100% real-world data using only a fraction (10-25%) of labeled real examples.
- Accurate simulation of physical consistency, including coherent environment backgrounds and realistic hand-object grasps, is critical for synthetic data's effectiveness, outperforming simpler data augmentation methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.