💡 This paper presents SeC, a concept-driven video object segmentation framework that progressively builds visual conceptual representations and only invokes large vision-language models at scene transit