Strategies to Leverage Foundational Model Knowledge in Object Affordance Grounding

Abstract

An important task for intelligent systems is affordance grounding, where the goal is to locate regions on an object where an action can be performed. Past weakly supervised approaches learn from human-object interaction (HOI) by transferring grounding knowledge from exocentric to ego-centric views of an object. The use of HOI priors is inherently noisy and thus provides a limited source of supervision. To address this challenge, we identify that recent foundational models (i.e. VLMs and LLMs) can serve as auxiliary sources of knowledge for frameworks due to their vast world knowledge. In this work, we propose strategies to extract and leverage foundational model knowledge related to attributes and object parts to enhance an HOI-based affordance grounding framework. In particular, we propose to combine HOI and foundational model priors through (1) a spatial consistency loss and (2) heatmap aggregation. Our strategies result in mKLD and mNSS improvements, and insights suggest future directions for improving affordance grounding capabilities.