TACO is a comprehensive dataset for studying bimanual hand-object interactions, supporting generalizable research in action recognition, motion forecasting, and grasp synthesis.
Abstract
Directory:
Introduction
Humans synchronize movements of both hands to manipulate objects.
Existing technical approaches are limited to single hand-object interactions.
Constructing TACO
TACO dataset spans various tool-action-object compositions for daily activities.
Data acquisition pipeline combines multi-view sensing with motion capture system.
Data Extraction System
Captures hand and object motions from multiple camera views.
Dataset Statistics Comparison
Compares TACO with existing 3D hand-object interaction datasets.
Experiments Overview:
Compositional Action Recognition: Evaluates model performance on different test sets.
Generalizable Hand-Object Motion Forecasting: Measures accuracy under various generalization settings.
Cooperative Grasp Synthesis: Assesses physical plausibility and realism of generated grasps.