뇌역공학과 영상연구실

A large-scale benchmark for reasoning segmentation that goes beyond the single-target, single-granularity setting most prior work assumed. MMR forces a model to ground multiple referents at multiple levels of granularity in the same scene, exposing a dimension where current vision–language models still struggle. Existing reasoning segmentation benchmarks ask the model to find one object given a query. Real scenes rarely come like that. A question can refer to several objects, or to a part vs. a whole vs. a category at the same time. MMR introduces ~194K samples covering multi-target and multi-granularity references across diverse images. We use the benchmark to evaluate state-of-the-art LMM segmenters and show systematic failure modes: models hallucinate extra masks under multi-target queries and confuse part / whole hierarchies. We also propose a baseline that explicitly decodes a variable number of masks at variable granularity. We work at the intersection of vision AI and large language models. We build systems that perceive, reason, and explain. A research lab at KAIST EE, focused on vision AI and language. The Brain Reverse Engineering & Imaging Lab (BREIL) is led by Prof. Dae-Shik Kim at the School of Electrical Engineering, KAIST. The lab grew out of computational neuroscience and brain imaging. Its current focus is the intersection of vision AI and large language models: building systems that perceive scenes, reason about them, and explain their decisions in language. Our work appears regularly at top venues including ICLR, ICCV, CVPR, WACV, BMVC, MICCAI, IEEE TIP, and IEEE TKDE. Beyond papers, the lab has produced books and registered patents that translate research into practice.