SingleOctoNet: TableADemonstrationLarge-Scale Multi-Modal(NeurIPS 2025DatasetFormat)for Human Activity Understanding Grounded in Motion-Captured 3DPose Labels

Activities are color-coded by category, revealing modality-specific recognition patterns. A.2 HPE task details Figure 3 demonstrates representative examples of ground truth versus predicted 3D human poses from the best-performing baseline model across different input modalities. The selected samples showcase diverse poses that effectively highlight model performance characteristics.