FMDL: Enhancing Open-World Object Detection with foundation models and dynamic learning

Yangyang Huang, Jie Hu, Ronghua Luo

Abstract: Highlights•We apply UDA to OWOD and propose DLDUA for dynamic learning.•We propose DON to reduce the domain gap in DLDUA, improving accuracy.•Foundational models generate pseudo-labels to enhance DLDUA performance.•Our method outperforms SOTA on the OWOD detection benchmark.