Intergen
Intergen
Intergen
2022.10 - 2023.4
Role
Reseach Assistant
Reseach Assistant
Awards and honor
Published on International Jounral of Computer Vision
Published on International Jounral of Computer Vision
Introduction
InterGen is a diffusion-based approach for generating realistic two-person interaction motions, built upon the large-scale multimodal dataset InterHuman (around 107 million frames and 16,756 text descriptions). By explicitly modeling global relations between two performers and introducing novel spatial regularization terms, InterGen generates more diverse and compelling two-person motions than prior methods, and supports various downstream applications involving multi-human interactions.
InterGen is a diffusion-based approach for generating realistic two-person interaction motions, built upon the large-scale multimodal dataset InterHuman (around 107 million frames and 16,756 text descriptions). By explicitly modeling global relations between two performers and introducing novel spatial regularization terms, InterGen generates more diverse and compelling two-person motions than prior methods, and supports various downstream applications involving multi-human interactions.
Contribution
Captured two-person motions with 76 synchronized cameras, leveraging Ground Truth for high-fidelity data
Built the InterHuman dataset (107 million frames, accurate skeletal motions, 16,756 descriptions)
Developed a tailored motion diffusion model with cooperative transformers, mutual attention, and novel motion input
Captured two-person motions with 76 synchronized cameras, leveraging Ground Truth for high-fidelity data
Built the InterHuman dataset (107 million frames, accurate skeletal motions, 16,756 descriptions)
Developed a tailored motion diffusion model with cooperative transformers, mutual attention, and novel motion input
Captured two-person motions with 76 synchronized cameras, leveraging Ground Truth for high-fidelity data
Built the InterHuman dataset (107 million frames, accurate skeletal motions, 16,756 descriptions)
Developed a tailored motion diffusion model with cooperative transformers, mutual attention, and novel motion input
