Bridging the Gap between Geometry and User Intent: Retrieval of CAD Models via Regions of Interest

Abstract

CAD model retrieval is a key functionality for increasing the efficiency of product design and development. A major challenge for content-based retrieval solutions is to bridge the semantic gap between geometric information and design intent. Current retrieval solutions lack capabilities to tailor search results to user interpretation and application context. This work presents an interactive retrieval system on heterogeneous CAD model collections that leverages user-defined regions of interest (ROIs) to express design intent in an intuitive manner. Our proposed system uses self-supervised learning techniques to capture nuanced semantic shape relationships and employs non-exhaustive search techniques to perform large-scale partial shape matching in real-time. We report significant performance increases over global retrieval on three mechanical component datasets. We also outline the scalability of the approach to large datasets and multi-modal data.