The paper introduces PRIMA, a method for recognizing fine-grained bodily behaviors in social settings without requiring labeled examples for each behavior. It focuses on actions like self-grooming, postural shifts, and backchanneling, which are often subtle but meaningful in social interaction.
The approach combines pose refinement with an inductive motion adapter to generalize to new behaviors in a zero-shot manner. This is positioned as a useful intermediate step for computational social signal analysis, where the relevant behaviors can vary across different social contexts.
Since the source is a single arXiv abstract, full architectural details and evaluation results are not described here. The core contribution is the pairing of pose refinement with an inductive adapter to recognize social bodily behaviors without prior task-specific examples.