PHASE: PHysically-grounded Abstract Social Events for Machine Social Perception

Abstract

The ability to perceive and reason about social interactions in the context of physical environments is core to human social intelligence and human-machine cooperation. However, no prior dataset or benchmark has systematically evaluated physically grounded perception of complex social interactions that go beyond short actions, such as high-fiving, or simple group activities, such as gathering. In this work, we create a dataset of physically-grounded abstract social events, PHASE, that resemble a wide range of real-life social interactions by including social concepts such as helping another agent. PHASE consists of 2D animations of pairs of agents moving in a continuous space generated procedurally using a physics engine and a hierarchical planner. Agents have a limited field of view, and can interact with multiple objects, in an environment that has multiple landmarks and obstacles. Using PHASE, we design a social recognition task and a social prediction task. PHASE is validated with human experiments demonstrating that humans perceive rich interactions in the social events, and that the simulated agents behave similarly to humans. As a baseline model, we introduce a Bayesian inverse planning approach, SIMPLE (SIMulation, Planning and Local Estimation), which outperforms state-of-the-art feed-forward neural networks. We hope that PHASE can serve as a difficult new challenge for developing new models that can recognize complex social interactions.

Relational Visual Information explains Human Social Inference: A Graph Neural Network model for Social Interaction Recognition

“…Two other models (Fan et al, 2019;Wu et al, 2019) like SocialGNN operate directly on graphs of scene entities, but they either don't preserve temporal information or require intermediate supervision at smaller time scales. As a result, they are not well suited to match human behavior on extended events like the ones tested here (Netanyahu et al, 2021).…”

“…For VisualRN-Rel, along with the entity features, we append a relational input feature (boolean vector denoting which edges are present) to match the binary edge information provided to the GNN. Finally, we also compared SocialGNN and the baselines to the performance of an Inverse Planning model, which currently achieves state-ofthe-art performance on this task (Netanyahu et al, 2021).…”

“…Some videos featured long time windows at the end with no movement, so we trimmed all the videos to end 2secs after all the entities stopped moving, resulting in 2.75-25sec long videos. For more details on the dataset see Netanyahu et al, 2021.…”

“…We also compared our model's performance on the animated videos (PHASE) with a generative inverse planning model called SIMPLE (Netanyahu et al, 2021). For each video, SIMPLE generates a hypothesis for the possible physical/social goals and relationships of the agents, simulates trajectories corresponding to these hypotheses, and compares these simulated trajectories to the observed trajectories of the entities in the video.…”

“…These actions are then executed via the physics engine to get the simulated trajectory. For more details refer to the model paper (Netanyahu et al, 2021). For comparison on the standard PHASE dataset, we got the predictions for all videos (except one video among the 400 where the model gave a NaN value) and calculated the performance accuracy for each bootstrap.…”