ST-MoE for Multi-Person Motion Prediction
Analysis
This paper addresses the limitations of existing multi-person motion prediction methods by proposing ST-MoE. It tackles the inflexibility of spatiotemporal representation and high computational costs. The use of specialized experts and bidirectional spatiotemporal Mamba is a key innovation, leading to improved accuracy, reduced parameters, and faster training.
Key Takeaways
- •Proposes ST-MoE to improve multi-person motion prediction.
- •Addresses limitations of existing methods: inflexible spatiotemporal representation and high computational cost.
- •Employs specialized spatiotemporal experts and bidirectional spatiotemporal Mamba.
- •Achieves state-of-the-art accuracy with reduced parameters and faster training.
Reference
“ST-MoE outperforms state-of-art in accuracy but also reduces model parameter by 41.38% and achieves a 3.6x speedup in training.”