A critical review of sketch collection methods: remembering how humans really sketch

Abstract

Sketching is a natural and effortless mode of expression that does not require specialized knowledge. As a powerful tool for processing and communicating ideas, sketching promises immense opportunities in the field of Human–Computer Interaction (HCI). However, with the increased need for data to train intelligent systems, sketch collection practices have shifted toward easy-to-collect data without sufficient consideration of how collection conditions shape what the resulting data represents — and whether it matches the intended downstream use. This paper addresses the ecological validity gap in sketch datasets through four contributions: (1) we compile a comprehensive list of factors that significantly affect the nature of sketch datasets; (2) we survey 86 sketch datasets and present a critical review with respect to these factors; (3) we critique dataset construction practices that produce data labeled as "sketches" without capturing authentic sketching behavior; and (4) we provide good practices for collecting sketches that reflect a broader range of human sketching behavior, and advocate for deliberate dataset selection and construction relative to downstream use. Overall, we argue that the community must return to collecting data that accurately reflects how people actually sketch, enabling the effective integration of sketching into HCI applications.