# DMMGAN: Diverse Multi Motion Prediction of 3D Human Joints using Attention-Based Generative Adverserial Network

**Authors:** Payam Nikdel, Mohammad Mahdavian, Mo Chen  
**arXiv:** [2209.09124](https://arxiv.org/abs/2209.09124)  
**Date:** 2022-09-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2209.09124v2/

## Key points

- Accurate human motion prediction is crucial in human-robot applications
- Existing studies oversimplify the problem by limiting models to predicting relative to a fixed joint or forecasting only one potential future motion
- Proposed transformer-based generative model aims to forecast multiple diverse human motions by analyzing historical data on human motion
- Approach first predicts body pose relative to the hip joint and employs a specialized "Hip Prediction Module" for forecasting hip movements
- Similarity loss function introduced to enhance diversity of predicted motions by penalizing pairwise sample distances
- Extensive experimentation shows that the system surpasses current state-of-the-art methods in human motion prediction, excelling in generating varied multi-motion trajectories while accurately capturing hip movements
- Research contributes significantly to advancing capabilities of human-robot interaction systems and enables more sophisticated applications in domains like healthcare, entertainment, and assistive technologies

## Summary

In the realm of human-robot applications, accurate human motion prediction is crucial. However, existing studies oversimplify the problem by limiting their models to predicting relative to a fixed joint or forecasting only one potential future motion. This approach fails to capture the complexity and diversity of human actions as a single output cannot encompass all possible movements. To address these limitations, this paper introduces a novel transformer-based generative model designed to forecast multiple diverse human motions. By leveraging this model, researchers aim to generate a range of future possibilities (denoted as \textit{N}) by analyzing historical data on human motion. The proposed approach first predicts the body pose relative to the hip joint and then employs a specialized "Hip Prediction Module" to forecast the trajectory of hip movements for each predicted pose frame. To enhance the diversity of predicted motions, a similarity loss function is introduced to penalize pairwise sample distances. Through extensive experimentation and evaluation, the authors demonstrate that their system surpasses current state-of-the-art methods in human motion prediction. Notably, their model excels in generating varied multi-motion trajectories while accurately capturing hip movements—a critical aspect often overlooked in existing predictive frameworks. By offering a more comprehensive and nuanced understanding of human motion dynamics, this research contributes significantly to advancing the capabilities of human-robot interaction systems and paves the way for more sophisticated applications in various domains such as healthcare, entertainment, and assistive technologies.