Core Concepts
Proposing ReGenNet for human action-reaction synthesis benchmark.
Abstract
ReGenNet introduces a novel approach to human action-reaction synthesis, focusing on asymmetric, dynamic, synchronous, and detailed interactions. The model generates instant and plausible human reactions conditioned on given actions. By annotating actor-reactor orders in datasets like NTU120, Chi3D, and InterHuman, ReGenNet achieves state-of-the-art results in FID scores, action recognition accuracy, diversity, and multi-modality. The model is modular and flexible for various settings of conditional action-reaction generation.
Stats
NTU120 dataset includes 8,118 interaction sequences with 26 action categories.
InterHuman dataset contains 6,022 interaction sequences captured by a motion capture studio.
Chi3D dataset consists of 373 interaction sequences for testing the model's generalization ability.
Quotes
"ReGenNet can generate instant and realistic reactions compared to baselines."
"Our contributions include analyzing asymmetric human-human interactions and proposing a benchmark for action-reaction synthesis."