Open-vocabulary semantic segmentation aims to assign pixel­level labels using arbitrary text queries, but existing CLIP-based methods often produce diffuse similarity maps and struggle with precise boundaries or small x-x-objects. Two-stage approaches―first generating class-agnostic mask proposals, then aligning them to text embeddings―improve shape fidelity but suffer from over-segmentation when using dense SAM proposals and from coarse boundaries when using patch­based spectral clustering. To address these limitations, we propose a novel graph-construction pipeline in which nodes correspond to SAM’s class-agnostic masks―providing fine boundary precision―and are augmented with textual nodes representing the target class embeddings. Edge weights capture both inter-mask visual similarity and structural affinity via cosine similarity in a joint embedding space, ensuring that clusters reflect full-x-x-object semantics rather than only discriminative parts. Our dual visual-structural graph promotes geometrically coherent, semantically complete mask proposals without the need for costly post-processing refinements. With the proposed method, state-of-the-art performance is attained over the latest approaches on several benchmarks. 

e-mail: choivy9102@pusan.ac.kr