<ul><li>Recent multi-teacher distillation methods have unified the encoders of multiple foundation models into a single encoder, achieving competitive performance on core vision tasks.</li><li>The paper introduces the concept of heterogeneous teacher distillation, where teacher models vary significantly in design objectives and the data they were trained on.</li><li>The researchers propose DUNE, a single encoder excelling in 2D vision, 3D understanding, and 3D human perception, achieving performance comparable or even surpassing larger teachers on their respective tasks.</li><li>DUNE outperforms MASt3R in Map-free Visual Relocalization with a smaller encoder.</li></ul>

DUNE: Distilling a Universal Encoder from Heterogeneous 2D and 3D Teachers

Discover more