SocialStructureHHI — Phase-Decomposition Dataset

Phase-decomposed two-person interaction data used to train SocialStructureHHI (Social Structure Matters in 3D Human–Human Interaction Generation). Derived from Inter-X (ix) and InterHuman (ih); each multi-person sequence is segmented into ordered phases, and each phase carries per-actor motion and per-actor text.

Files

file	content
socialstructure_hhi_motion_train.h5	per phase-actor motion samples (train)
socialstructure_hhi_motion_val.h5	per phase-actor motion samples (val)
socialstructure_hhi_motion_test.h5	per phase-actor motion samples (test)
socialstructure_hhi_text.h5	per-sequence phase decomposition (text)
motion_norm_stats.npz	per-dim mean/std over the train split (201-dim)

Outlier phases (>300 frames) are already excluded; only valid samples are packed.

Motion .h5 layout

/<phase_id>_<actor>            e.g. "IH_1008_phase000_P1"
  attrs: sequence_id, phase_id, actor_id, dataset {ix|ih}, phase_type,
         interaction_category, phase_len_bucket, t_start, t_end,
         phase_text, global_text
  datasets:
    target_motion_local    [W, 201]   the actor's motion for this phase
    seq_mask               [W]        True = real frame, False = padding
    contact_frame_mask     [W]        True = contact frame
    prefix_mask            [W]        True = read-only overlap (inpaint) frame
    self_history_local     [10, 201]  actor's 10-frame history prefix
    partner_history_local  [10, 201]  partner's 10-frame history
    partner_prev_local     [W, 201]   partner motion, previous phase (ego frame)
    partner_curr_local     [W, 201]   partner motion, current phase (ego frame)
    partner_prev_seq_mask  [W]
    partner_curr_seq_mask  [W]
    partner_prev_time_ids  [W+10]
    partner_curr_time_ids  [W+10]
    initial_frame          [201]      actor's first frame (ego frame)
    partner_initial_frame  [201]      partner's first frame, in actor ego frame

Motion layout (201-dim/frame): root translation (3) + root 6D rotation (6) + 21 body-joint 6D rotations (126) + 22 forward-kinematics joint positions, root-relative (66). W is the per-phase duration bucket ∈ {30,60,90,120,150,180,240,300} @ 30 fps. All motion is in the actor's ego-centric frame.

Text .h5 layout

/<split>/<seq_id>
  attrs: dataset {ix|ih}, global_text
  /<phase_key>                 e.g. "phase000"
    attrs: p1_action, p2_action, phase_type
    dataset: frame_range [2]   (start, end) in the source sequence

Loading

import h5py, numpy as np
with h5py.File("socialstructure_hhi_motion_val.h5") as f:
    g = f["IH_1008_phase000_P1"]
    motion = g["target_motion_local"][:]     # [W, 201]
    text   = g.attrs["phase_text"]
    print(g.attrs["phase_type"], motion.shape, text)

stats = np.load("motion_norm_stats.npz")     # mean[201], std[201]

Links

• 📄 Paper: https://arxiv.org/abs/2606.24255
• 💻 Code: https://github.com/EngineeringAI-LAB/SocialStructureHHI
• 🌐 Project page: https://engineeringai-lab.github.io/SocialStructureHHI/
• 🤗 Model: https://huggingface.co/EngineeringAI-LAB/SocialStructure

Source datasets & citation

Built on Inter-X and InterHuman; please also cite the original datasets.

@misc{wang2026socialstructure,
  title  = {Social Structure Matters in 3D Human--Human Interaction Generation},
  author = {Zhongju Wang and Beier Wang and Yatao Bian and Pichao Wang and Zhi Wang
            and Daoyi Dong and Hongdong Li and Huadong Mo and Zhenhong Sun},
  year   = {2026},
  eprint = {2606.24255},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}