Achieve pixel-level grounding in LMMs without explicit grounding supervision.

Learn large multimodal models with RLHF augmented by factual information to reduce hallucination.

Learn to segment visual entities and their parts in an open-world with pure self-supervision.

Develop a fully self-supervised approach to object detection which understands part-to-whole composition and needs much less training costs.

Distill knowledge from an automatically designed sequence of teachers into a lightweight student object detector to mitigate the teacher-student capacity gap.

Unify contrastive learning for representation learning and Mean Teacher for domain adaptation into one general-purpose framework for learning unsupervised domain adaptive object detectors.

Analyze the slow convergence of DEtection TRansformer (DETR), and propose alternative solutions to improve convergence and performance of DETR.