Microsoft Mixed Reality & AI Lab Cambridge · Visual Understanding

DAViD

Data-efficient vision model trained exclusively on 300K high-fidelity synthetic images that achieves state-of-the-art accuracy on dense prediction tasks (depth estimation, surface normal estimation, soft foreground segmentation) while requiring significantly lower training and inference costs than foundational models. Demonstrates that procedurally-generated synthetic data with perfect labels can match or exceed real-data training performance while enabling explicit control over data diversity and fairness.