Paper #30: SEA: Understanding Sketch Abstraction Efficiency via Element-Level Commonsense for Human-AI Sketching

Sketches communicate ideas through compact visual abstractions, yet existing methods based on reference images, low-level visual features, or recognition accuracy provide limited insight into this defining property of sketches. We introduce SEA ($\textbf{S}$ketch $\textbf{E}$valuation metric for $\textbf{A}$bstraction efficiency), a reference-free metric that measures how economically a sketch preserves class-defining visual elements while maintaining semantic recognizability. SEA derives these elements from commonsense knowledge about features typically depicted in sketches and uses visual question answering (VQA) to estimate their presence under visual economy. We also introduce CommonSketch, a semantically annotated dataset of 23,100 human-drawn sketches across 300 classes, each paired with captions and element-level annotations. Experiments show that SEA aligns with human judgments and distinguishes abstraction efficiency across sketching conditions, while CommonSketch enables fine-grained evaluation of sketch understanding in vision-language models (VLMs).