InterMamba: Efficient Human-Human Interaction Generation With Adaptive Spatio-Temporal Mamba

Area of Science:

• Computer Vision
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Human-human interaction generation is crucial for understanding social behavior and is a key area in motion synthesis.
• Existing transformer-based methods struggle with scalability and efficiency in generating complex human interactions.
• There is a need for advanced models that can effectively capture long-range dependencies in motion data.

Purpose of the Study:

• To propose InterMamba, a novel and efficient method for human-human interaction generation.
• To leverage the Mamba framework for improved capture of long-sequence dependencies and real-time feedback.
• To enhance the quality and efficiency of motion synthesis for social interactions.

Main Methods:

• Introduced an adaptive spatio-temporal Mamba framework with parallel SSM branches for integrating spatial and temporal motion features.
• Developed self and cross adaptive spatio-temporal Mamba modules to capture intra- and inter-sequence motion dependencies.
• Utilized the Mamba architecture for efficient processing of long motion sequences.

Main Results:

• Achieved state-of-the-art performance on two human-human interaction datasets.
• Demonstrated remarkable quality and efficiency in generated motion sequences.
• Reduced model parameter size to 66M (36% of baseline InterGen) and achieved 46% of InterGen's inference time.

Conclusions:

• InterMamba offers a significant improvement over existing methods in human-human interaction generation.
• The Mamba-based framework provides a scalable and efficient solution for complex motion synthesis.
• The proposed method enables high-quality, real-time generation of human social interactions.