Paper #33: Object-Level Access to SVG Scenes: A Benchmark and Raster-Grounded Method for Text-Guided Vector Object Extraction

SVG graphics are editable structured documents, but semantic objects in SVG scenes often do not align with paths, groups, or other elements in the document tree. This makes object-level access difficult for creative workflows that require extracting, editing, or reusing parts of existing vector assets. We introduce \textit{text-guided vector object extraction}, where the goal is to recover an editable sub-SVG corresponding to a natural-language query. We contribute a benchmark with node-level annotations and vector-aware evaluation, containing 2,793 evaluation samples. We also propose a raster-grounded extraction method that localizes the queried object in the rendered image and maps the resulting region back to SVG primitives through overlap-based filtering. Experiments show that the proposed method provides a strong baseline for the benchmark and outperforms structured LLM baselines on node-level overlap metrics. Our results highlight the need for SVG-aware models and evaluation protocols for editable object-level vector graphics.