Minimize the synthetic-to-real domain gap

Develop and assess methods that minimize the performance gap between synthetic and real data for egocentric hand–object interaction detection, such as domain adaptation strategies or targeted synthetic data improvements.

Background

Reducing the gap is a core objective for making synthetic data practically useful in real-world deployments. The paper evaluates multiple strategies, including unsupervised and semi-supervised domain adaptation and alignment of objects, grasps, and environments.

A systematic approach to minimizing the gap enables effective use of synthetic data when real annotations are scarce or unavailable.