InterAct Dataset

InterAct is a multi-modal two-person interaction dataset for research in human motion, facial expressions, and speech. For details, please refer to our webpage.

Quick Start

A Quick Start Jupyter notebook is provided at quickstart.ipynb. It covers examples for:

1. Querying the scenario and actor databases
2. Finding actor pairs for a recording session
3. Loading performance data (BVH, face parameters, audio)
4. Visualizing face blendshapes over time
5. Loading both actors in a two-person interaction

Repository Structure

Database Files

scenarios.db

SQLite database containing scenario metadata with the following tables:

• scenarios: Contains scenario definitions

  • id (INTEGER): Scenario ID (used in filenames)
  • relationship_id (INTEGER): FK to relationships table
  • primary_emotion_id (INTEGER): FK to emotions table
  • character_setup (TEXT): Character context description
  • scenario (TEXT): Scenario description

• relationships: Relationship types between actors (e.g., "architect / contractor", "boss / subordinate")

  • id (INTEGER): Relationship ID
  • name (VARCHAR): Relationship description

• emotions: Primary emotion categories (e.g., "admiration", "anger", "amusement")

  • id (INTEGER): Emotion ID
  • name (VARCHAR): Emotion name

actors.db

SQLite database containing actor and session information:

• actors: Actor metadata

  • actor_id (TEXT): Three-digit actor ID (e.g., "001", "002")
  • gender (TEXT): "male" or "female"

• sessions: Recording session information

  • date (TEXT): Session date in YYYYMMDD format
  • male_id (TEXT): Actor ID of the male participant
  • female_id (TEXT): Actor ID of the female participant

---

Data Directories

Motion and facial data are provided here at 30 fps. The performance data files follow this naming convention:

<date>_<actor_id>_<scenario_id>.<extension>

Example: 20231119_001_051.bvh = recorded on 2023-11-19, actor 001, scenario 51

bvhs/

BVH motion capture files of the performances.

bvhs_retarget/

Retargeted BVH files for use in body_to_render.blend.

face_ict/

Facial blendshape parameters in ICT-FaceKit format (shape: (N, 55)). Suitable for training models and rendering with face_ict_to_render.blend.

face_arkit/

Facial blendshape parameters in ARKit format (shape: (N, 51)). Used in body_to_render.blend for full body visualization.

face_ict_templates/

Base mesh templates in ICT-FaceKit topology, named by actor ID (e.g., 001.obj). Useful for training models.

wav/

Audio recordings from each actor in each performance.

body_renders/

Pre-rendered full-body visualizations (body + face + audio) as MP4 videos. These files use a different naming convention since they contain both actors:

<date>_<scenario_id>.mp4

Example: 20231119_051.mp4 = scenario 51 recorded on 2023-11-19

lip_acc/

Additional 1-hour facial dataset with attention to accuracy of lip shapes and pronunciation. Only one actor (006) was captured in this dataset, and the scenario_id of these files correspond to the order of the sentences in lip_acc_sentences.txt. Useful for fine-tuning.

---

Scripts (scripts/)

Blender Files

• body_to_render.blend: Blender project for rendering full-body (face+body) visualizations. Contains pre-configured character rigs mapped to actor IDs. The "composite scene in dataset" script reads job files, composites both actors with BVH body motion from bvhs_retarget/ and ARKit face blendshapes from face_arkit/. The "render all scenes" script renders MKV videos to body_renders_noaudio/.

• face_ict_to_render.blend: Blender project for rendering face-only visualizations using ICT-FaceKit topology. Contains pre-configured actor mesh scenes (mesh-001, mesh-002, etc.) and a "composite scenes and render" script that reads job files, loads blendshape animations from face_ict/, and renders 1080x1080 PNG sequences at 30fps using EEVEE. Output goes to face_renders_noaudio/.

Conversion Scripts

• face_ict_to_arkit.py: Converts ICT-FaceKit blendshape parameters (55 blendshapes) to ARKit format (51 blendshapes). Merges certain blendshape pairs and removes unused indices.

• face_ict_to_vertices.py: Converts ICT blendshape parameters to vertex sequences using the blendshape basis matrix. Outputs per-frame vertex positions as numpy arrays with shape (N, V*3), where coordinates are packed contiguously per vertex: [v1x, v1y, v1z, v2x, v2y, v2z, ...].

Render Utilities

• render_add_audio.py: Combines rendered video with audio tracks. Supports both face renders (single actor) and body renders (mixed audio from both actors).

Data Files

• blendshape_ict.npy: ICT-FaceKit blendshape basis matrix used for converting blendshape parameters to vertex offsets, used in face_ict_to_vertices.py.

Job Files

We recommend using a job file and splitting the rendering into batches, as opposed to rendering all scenarios in one go.

• example_body_render_job.txt: Example job file listing scenes to render in body format (<date>_<scenario_id>).
• example_face_render_job.txt: Example job file listing scenes to render in face format (<date>_<actor_id>_<scenario_id>).

Errata

• The face files for 20240126_006_034 is unavailable due to a conversion issue. When rendering the scene in body_to_render.blend, the female face blendshape animations are not applied.

Acknowledgements

body_to_render.blend is based on the visualization Blender project kindly provided by the BEAT dataset authors.

If you used InterAct as part of your research, please cite as following:

@article{ho2025interact,
  title={InterAct: A Large-Scale Dataset of Dynamic, Expressive and Interactive Activities between Two People in Daily Scenarios},
  author={Ho, Leo and Huang, Yinghao and Qin, Dafei and Shi, Mingyi and Tse, Wangpok and Liu, Wei and Yamagishi, Junichi and Komura, Taku},
  journal={Proceedings of the ACM on Computer Graphics and Interactive Techniques},
  volume={8},
  number={4},
  pages={1--27},
  year={2025},
  publisher={ACM New York, NY},
  doi={10.1145/3747871}
}