InterPet4D (v1.1)

Authors: Yichen Peng*, Jyun-Ting Song*, Chen-Chieh Liao*, Kris Kitani, Hideki Koike, Erwin Wu _{*Equal contribution.}

InterPet4D is a multimodal, ego-centric dataset of natural human–pet (dog) interactions. Each clip provides time-synchronized audio, SMPL-X human body motion, MANO hand motion, pet skeletal motion, and SMAL pet body parameters, enabling research on cross-species interaction, multimodal motion generation, audio-conditioned animation, and animal behavior understanding.

Highlights

• 228 English clip-level interaction captions, available as per-clip TXT and aggregate JSON/CSV.

• 113 interaction sessions across 13 dogs (dog00–dog12) and ~23 participants (p01–p23).

• 227 ego-centric clips (~17–20 s each) captured with head-mounted glasses (Project Aria–style).

• Time-aligned modalities per clip: raw audio, MERT audio embeddings, SMPL-X body, MANO hands, pet skeleton, and SMAL pet body fits.

• Cleaned SMAL motion release for 225 clips, with direct 6D rotations, temporal cleaning, world-up coordinates, initial-yaw normalization, and a fixed zero-height floor.

Dataset Structure

interpet4d_ver1/
├── video_caption/       # English clip captions (TXT, JSON, CSV)
├── interpet_audio/      # Raw audio (.mp3)
├── interpet_mert/       # MERT pre-extracted audio embeddings (.npy)
├── smpl_npy/            # SMPL-X human body parameters (.npy, dict)
├── mano_npy/            # MANO left/right hand parameters (.npy, dict)
├── pet_npy/             # Pet (dog) 3D keypoint trajectories (.npy)
├── smal_npy/            # Raw SMAL pet body fits (.npz, stacked per-frame)
└── cleanup_smal/         # Cleaned/canonicalized SMAL motion
    ├── manifest.json
    ├── stats.npz
    └── clips/             # One packed .npz per available clip

File Naming Convention

interpet_dog{DD}_p{PP}_take{TT}_ego_{NNN}.{mp3|npy}
              │      │       │        └── clip index within the take
              │      │       └────────── take number
              │      └────────────────── participant ID
              └───────────────────────── dog ID

The basename (without extension) is the clip ID, shared across all directories — use it as the join key.

Modality Specifications

Folder	Format	Shape / Schema	Notes
interpet_audio/	MP3	48 kHz, stereo	Ego-microphone audio.
interpet_mert/	.npy	(T_a, 1024) float32	MERT features at ~75 Hz.
smpl_npy/	.npy (dict)	see below	Per-subject SMPL-X parameters.
mano_npy/	.npy (dict)	see below	{'left': {...}, 'right': {...}}.
pet_npy/	.npy	(T_p, 20, 4) float32	20-joint pet skeleton; last axis is (x, y, z, score).
smal_npy/	.npz	see below	SMAL pet body fits, per-frame parameters stacked over time.
cleanup_smal/clips/	.npz	see below	Cleaned direct-6D SMAL motion at 30 Hz, packed per clip.

SMPL-X dict schema (key = subject id, e.g. aria01):

{
  'global_orient': (T, 3)        # axis-angle root orientation
  'transl':        (T, 3)        # root translation (meters)
  'body_pose':     (T, 69)       # 23 joints × 3 (axis-angle)
  'betas':         (T, 10)       # SMPL-X shape coefficients
  'joints':        (T, 45, 3)    # 3D joint positions
  'vertices':      (T, 6890, 3)  # exported body mesh vertices
  'epoch_loss':    (T,)          # optimization residual
}

MANO dict schema ('left' / 'right', each):

{
  'joints': (T, 1, 21, 3)        # 21 hand keypoints (3D)
  'pose':   (T, 16, 3, 3)        # rotation matrices for 16 joints
  'transl': (T, 3)               # wrist translation
}

SMAL (smal_npy/) schema — per-clip .npz with all frames stacked:

{
  'pose_rotmat':  (T, 35, 3, 3)  # SMAL joint rotations (rotation matrices)
  'betas':        (T, 30)        # SMAL shape coefficients
  'betas_limbs':  (T, 7)         # limb-specific shape coefficients
  'R_world':      (T, 3, 3)      # global rotation in world frame
  't_world':      (T, 3)         # global translation in world frame (meters)
  's_world':      (T,)           # global scale
  'kp_world':     (T, 24, 3)     # 24 keypoints in world coordinates
  'kp_weight':    (T, 24)        # per-keypoint confidence weight
  'frame_idx':    (T,) int32     # original frame index (sparse / non-contiguous)
}

SMAL fits cover 226 of 227 clips (one clip lacks fits). Frame indices in frame_idx are not necessarily contiguous — use them to align with the raw video frame rate.

Cleaned SMAL (cleanup_smal/) schema — 225 clips and 137,710 frames:

{
  'feature':                    (T, 213)  # concatenated motion feature
  'local_rotation_6d':          (T, 204)  # 34 local joints × 6D
  'root_rotation_6d':           (T, 6)
  'root_translation':           (T, 3)    # dog-body units
  'frame_idx':                  (T,)      # original sparse frame index
  'segments':                   (S, 2)    # valid [start, end) ranges
  'temporal_jump_after':        (T - 1,)
  'shape':                      (37,)     # betas + betas_limbs
  'betas':                      (30,)
  'betas_limbs':                (7,)
  'fixed_scale':                ()
  'initial_world_rotation':     (3, 3)
  'initial_world_translation':  (3,)
  'ground_world_z':             ()
  'ground_source':              ()
  'reprojection_px':            (T,)
  'quality_weight':             (T,)
}

The 213D feature layout is [local_rotation_6d (204), root_rotation_6d (6), root_translation (3)]. Local poses passed through the frozen BARC v3 Real-NVP representation and were decoded back to direct rotations before temporal quaternion/translation smoothing. Root motion is world-up, aligned to the initial yaw, expressed relative to the initial horizontal position, and floor-normalized to z = 0. This is a cleaned and canonicalized release, not an image-conditioned SMAL refinement result.

cleanup_smal/clips/ uses clip IDs without the leading interpet_ prefix. For example, dog01_p01_take01_ego_001.npz corresponds to interpet_dog01_p01_take01_ego_001 in the other modality folders. The two clips excluded from this release are dog01_p01_take01_ego_002 (no SMAL fit) and dog06_p14_take01_ego_001 (non-finite shape parameters).

Note on temporal alignment. All modalities are aligned by clip ID. Body / hand / pet motion are sampled at the same frame rate T; MERT features are at a higher rate T_a. Resample with the clip duration when fusing.

Loading Example

import numpy as np
import librosa

clip_id = "interpet_dog01_p01_take01_ego_001"

audio, sr  = librosa.load(f"interpet_audio/{clip_id}.mp3", sr=None)
mert       = np.load(f"interpet_mert/{clip_id}.npy")             # (T_a, 1024)
pet        = np.load(f"pet_npy/{clip_id}.npy")                   # (T, 20, 4)
mano       = np.load(f"mano_npy/{clip_id}.npy",  allow_pickle=True).item()
smplx      = np.load(f"smpl_npy/{clip_id}.npy",  allow_pickle=True).item()
smal       = np.load(f"smal_npy/{clip_id}.npz")                  # dict-like
cleanup_id = clip_id.removeprefix("interpet_")
clean_smal = np.load(f"cleanup_smal/clips/{cleanup_id}.npz")

print(audio.shape, sr)
print(smplx['aria01']['body_pose'].shape)
print(mano['right']['joints'].shape)
print(smal['pose_rotmat'].shape, smal['frame_idx'][:5])
print(clean_smal['feature'].shape, clean_smal['frame_idx'][:5])

Or via the datasets library:

from datasets import load_dataset
ds = load_dataset("ohicarip/interpet4d")

Video Captions

This caption release contains 228 English clip-level descriptions. Use recording_id to join captions to available audio and motion files; modality coverage can differ, so intersect the IDs when building paired datasets. Of these, 227 match the 227 published audio clips. Caption IDs without a published audio file: interpet_dog08_p17_take03_ego_001.

Captions describe the observed human–dog interaction and its sequence of actions. The annotation records report review of sampled multi-view video at 0.5-second intervals and Whisper large-v3 ASR transcripts. All 228 records have status="reviewed_video_and_asr", audio_transcribed=true, and transcript_reviewed=true; audio_reviewed=false denotes that direct audio-listening review was not performed. Expected source-audio languages are recorded per clip; the captions themselves are in English. ASR can misrecognize commands, so consult review_notes when interpreting descriptions of spoken instructions. These are clip-level summaries without sentence-level timestamps.

Files and fields

File	Contents
video_caption/<recording_id>.txt	One UTF-8 caption per clip.
video_caption/captions.json	Object keyed by recording_id, preserving typed annotation metadata.
video_caption/captions.csv	One row per clip with the same fields, for tabular loading.

Field	Meaning
recording_id	Full clip ID, including the interpet_ prefix; join key for other modalities.
caption	English interaction description.
duration_seconds	Duration of the source video in seconds; encoded audio duration may differ slightly.
evidence, status	Annotation evidence and review status.
expected_audio_language	Expected source-audio language code.
audio_reviewed, audio_transcribed, transcript_reviewed	Direct-listening, ASR, and transcript-review flags.
visual_sample_interval_seconds	Interval between sampled visual frames.
source_video, transcript_file, visual_review_pages	Provenance paths from the annotation workspace.
review_notes	Clip-specific interpretation notes, when present.

The provenance paths refer to the original annotation workspace. The referenced source previews, transcripts, and review images are not included in this caption release and those paths are not download links. CSV stores flags and list-valued metadata as text; use JSON to retain their original types.

Loading captions

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="ohicarip/interpet4d",
    repo_type="dataset",
    filename="video_caption/captions.json",
)
with open(path, encoding="utf-8") as f:
    captions = json.load(f)

clip_id = "interpet_dog01_p01_take01_ego_001"
print(captions[clip_id]["caption"])
# For cleaned SMAL motion, remove the leading interpet_ prefix:
cleanup_id = clip_id.removeprefix("interpet_")

For a tabular dataset, load the caption CSV explicitly:

from datasets import load_dataset

captions_ds = load_dataset(
    "csv",
    data_files="hf://datasets/ohicarip/interpet4d/video_caption/captions.csv",
    split="train",
)

Captions support text-conditioned motion generation, interaction retrieval, and multimodal alignment, and are distributed under the dataset's CC BY-NC 4.0 license.

SMAL Releases

• smal_npy/ contains the raw SMAL fits from the automated fitting pipeline and covers 226 clips.
• cleanup_smal/ contains the cleaned and canonicalized direct-6D motion representation and covers 225 clips. See cleanup_smal/manifest.json for split, segment, coordinate-system, and normalization metadata.

Intended Uses

• Cross-species (human ↔ dog) interaction modeling
• Audio-conditioned motion synthesis / vocal-to-motion translation
• Multimodal representation learning for animal behavior
• 4D scene understanding from ego-centric recordings

Ethical Considerations

• All participants provided informed consent for data release.
• No personally identifying information (faces / voices of bystanders) is included.
• Pet welfare: all interactions were supervised and non-coercive.

License

Released under CC BY-NC 4.0 — research and non-commercial use only. Commercial use requires explicit permission from the authors.

Citation

If you use InterPet4D in your research, please cite:

@dataset{interpet4d_2026,
  title  = {InterPet4D: A Multimodal Ego-Centric Dataset of Human--Pet Interactions},
  author = {Peng, Yichen and Song, Jyun-Ting and Liao, Chen-Chieh and Kitani, Kris and Koike, Hideki and Wu, Erwin},
  year   = {2026},
  url    = {https://huggingface.co/datasets/ohicarip/interpet4d},
  note   = {Version 1}
}

Changelog

• 2026-10-07 — Added video_caption/: 228 English interaction captions, annotation metadata, and caption loading examples.

• v1.1 (2026-10) — Added cleanup_smal/ and corrected the human body model name from SMPL to SMPL-X.

• v1 (2026-06) — Initial release: 227 clips, 13 dogs, ~23 participants, four time-aligned modalities.

Contact

For questions or commercial-use inquiries, please open a discussion on the Hugging Face repo or contact the authors directly.