Seeing is Knowing: Advancing Semantic Understanding with MLLMs in Grounding Tasks

Qihua Dong, Yue Bai, Yizhou Wang, Yun Fu

Keywords: Zero-shot segmentation; Multimodal LLM; MLLM Grounding

Abstract:

Large vision models (VLMs) achieve success in most daily scenarios but face challenges in special grounding tasks. This limitation is primarily due to insufficient semantic understanding for both tasks and images in current vision models. In contrast, large multimodal language models (M-LLMs) excel in semantic comprehension and instruction-following but underperform in detailed recognition. To harness the strengths of both, we propose to utilize M-LLMs to assist VLMs in handling difficult segmentation tasks.
The key to our approach involves
(1)leveraging M-LLMs for semantic expertise
and (2)formatting instruction-based guidance.
Our proposed framework is generalizable, performing well across various tasks. Experimental results show a significant performance improvement (10\%+) in challenging tasks like camouflage object detection, anomaly detection and medical image segmentation compared to zero-shot baselines.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8251