Abstract
Predicting long-term 3D human motion in social contexts remains a significant challenge due to the complex interdependencies between individuals and the tendency of traditional models to "freeze" or drift into implausible states over time. We present Social Diffusion (SDM), a generative framework that leverages diffusion models and social-aware aggregation to forecast realistic, multi-person interactions over extended durations (10+ seconds). By modeling motion as a denoising process conditioned on past observations, SDM maintains physical realism and social consistency, such as conversational turn-taking and role preservation, significantly outperforming existing state-of-the-art methods.
Introduction
Human motion is inherently social. Whether navigating a crowded room or engaging in a negotiation, our movements are continuously influenced by the actions and perceived intentions of others. While single-person motion prediction has seen substantial progress, multi-person anticipation often fails to capture these "social" nuances. Existing auto-regressive models frequently suffer from error accumulation, leading to static "mean poses" or socially nonsensical behaviors after just a few seconds.
Social Diffusion addresses these hurdles by moving away from step-by-step frame prediction. Instead, it treats motion anticipation as a generative task, producing entire sequences that respect the social roles and interaction patterns established in the observed past.
Methodology: Generative Denoising and Social Aggregation
The core of the Social Diffusion model is a Generative Denoising process. Unlike auto-regressive models that predict frame $N+1$ from frame $N$, SDM starts with Gaussian noise representing the entire future sequence and iteratively refines it into a coherent 3D motion path. This process is conditioned on a "prefix" (the observed past motion of all participants).
To ensure that the generated motions are interdependent, the architecture incorporates a Social Aggregation function ($\Gamma$). This order-invariant component allows information to flow between all individuals in the scene during the denoising steps. The researchers found that even simple averaging across individual features provides a powerful inductive bias for maintaining social consistency. Furthermore, the fully convolutional nature of the temporal networks allows the model to scale "zero-shot" to larger groups than those seen during training.
Experimental Evaluation
The model was evaluated on standard benchmarks (MuPoTS-3D, 3DPW, CMU-Mocap) and a specialized Haggling Dataset. The latter features triadic interactions between a buyer and two sellers, providing a rigorous test for social role preservation.
Metrics for Social Realism
Standard metrics like Mean Per Joint Positional Error (MPJPE) often fail to quantify the "social" quality of a prediction. To address this, the authors introduced the Symbolic Social Cues (SSCP) Protocol. This protocol maps continuous motion into discrete social states (e.g., "speaking" vs. "listening"). By comparing the distribution of these states in the predicted future to ground truth data, the researchers could verify the model's "social logic."
Key Results
Social Diffusion demonstrates remarkable stability and accuracy over long horizons:
- Accuracy: On the 3DPW dataset, SDM reduced MPJPE by over 54% for 3-second forecasts compared to previous state-of-the-art (SOTA) methods.
- Long-Term Stability: While SOTA models often degrade after 4 seconds, SDM maintains high Normalized Directional Motion Similarity (NDMS) even 10 seconds into the future.
- Role Preservation: In the Haggling dataset, SDM successfully "remembered" social roles. A buyer-detection network could identify the buyer in SDM's forecasts with ~89% accuracy, compared to near-random performance (~32%) from competing models.
Discussion and Future Work
Social Diffusion represents a significant step toward socially-aware AI. By prioritizing the global structure of interaction over local frame-to-frame transitions, it avoids the "freeze" common in previous models.
However, limitations remain. The model currently struggles with absolute global positioning over very long durations and assumes a fixed number of participants. Future research will focus on "open-world" scenarios where individuals join or leave groups dynamically and integrating environmental constraints like furniture to enable more robust applications in social robotics and virtual reality.
Summary updated