• Developed a multimodal benchmark to evaluate visual grounding capabilities in large vision-language models (LVLMs). Proposed a framework utilizing a large language model (LLM), such as GPT-4o, to enhance reasoning and object detection in LVLMs for improved visual grounding.• Proposed NIDS-Net, a few-shot novel instance detection and segmentation framework that detects objects by adapting vision foundation models and a Multimodal Large Language Model (MLLM). Proposed a Weight Adapter to align object-level semantics between templates and queries. Outperformed state-of-the-art detection methods by 22.3 and 46.2 AP. Ranked #1 on the world-wide BOP leaderboard. (Under review)• Designed a novel robotic system for improving unseen object instance segmentation by multi-object tracking and video object segmentation via long-term robot interaction, published in RSS 2023. Got media coverage• Improved real-world image segmentation and achieved state-of-the-art results through a designed transfer learning strategy. Demonstrated remarkable Boundary F-measure improvement from 65.1 to 89.3 on real-world images• Proposed a new vision Transformer architecture to combine feature representation and clustering for instance segmentation, achieving state-of-the-art performance: 87.3 Boundary F-measure on RGB-D OCID dataset and 21.1% improvement on RGB images. Published in ICRA 2024• Elevated generic deep neural network frameworks through log-supermodular CRFs, validated across diverse CV tasks. E.g., achieved a significant 9.7%enhancement in stereo matching. Published in NeurIPS 2022