Metadata-Version: 2.2 Name: seamless-interaction Version: 0.1.0 Summary: Library for seamless interaction dataset Author: Meta AI | FAIR Seamless Team License: Attribution-NonCommercial 4.0 International ======================================================================= Creative Commons Corporation ("Creative Commons") is not a law firm and does not provide legal services or legal advice. Distribution of Creative Commons public licenses does not create a lawyer-client or other relationship. Creative Commons makes its licenses and related information available on an "as-is" basis. Creative Commons gives no warranties regarding its licenses, any material licensed under their terms and conditions, or any related information. Creative Commons disclaims all liability for damages resulting from their use to the fullest extent possible. Using Creative Commons Public Licenses Creative Commons public licenses provide a standard set of terms and conditions that creators and other rights holders may use to share original works of authorship and other material subject to copyright and certain other rights specified in the public license below. The following considerations are for informational purposes only, are not exhaustive, and do not form part of our licenses. Considerations for licensors: Our public licenses are intended for use by those authorized to give the public permission to use material in ways otherwise restricted by copyright and certain other rights. Our licenses are irrevocable. Licensors should read and understand the terms and conditions of the license they choose before applying it. Licensors should also secure all rights necessary before applying our licenses so that the public can reuse the material as expected. Licensors should clearly mark any material not subject to the license. This includes other CC- licensed material, or material used under an exception or limitation to copyright. More considerations for licensors: wiki.creativecommons.org/Considerations_for_licensors Considerations for the public: By using one of our public licenses, a licensor grants the public permission to use the licensed material under specified terms and conditions. If the licensor's permission is not necessary for any reason--for example, because of any applicable exception or limitation to copyright--then that use is not regulated by the license. Our licenses grant only permissions under copyright and certain other rights that a licensor has authority to grant. Use of the licensed material may still be restricted for other reasons, including because others have copyright or other rights in the material. A licensor may make special requests, such as asking that all changes be marked or described. Although not required by our licenses, you are encouraged to respect those requests where reasonable. More considerations for the public: wiki.creativecommons.org/Considerations_for_licensees ======================================================================= Creative Commons Attribution-NonCommercial 4.0 International Public License By exercising the Licensed Rights (defined below), You accept and agree to be bound by the terms and conditions of this Creative Commons Attribution-NonCommercial 4.0 International Public License ("Public License"). To the extent this Public License may be interpreted as a contract, You are granted the Licensed Rights in consideration of Your acceptance of these terms and conditions, and the Licensor grants You such rights in consideration of benefits the Licensor receives from making the Licensed Material available under these terms and conditions. Section 1 -- Definitions. a. Adapted Material means material subject to Copyright and Similar Rights that is derived from or based upon the Licensed Material and in which the Licensed Material is translated, altered, arranged, transformed, or otherwise modified in a manner requiring permission under the Copyright and Similar Rights held by the Licensor. For purposes of this Public License, where the Licensed Material is a musical work, performance, or sound recording, Adapted Material is always produced where the Licensed Material is synched in timed relation with a moving image. b. Adapter's License means the license You apply to Your Copyright and Similar Rights in Your contributions to Adapted Material in accordance with the terms and conditions of this Public License. c. Copyright and Similar Rights means copyright and/or similar rights closely related to copyright including, without limitation, performance, broadcast, sound recording, and Sui Generis Database Rights, without regard to how the rights are labeled or categorized. For purposes of this Public License, the rights specified in Section 2(b)(1)-(2) are not Copyright and Similar Rights. d. Effective Technological Measures means those measures that, in the absence of proper authority, may not be circumvented under laws fulfilling obligations under Article 11 of the WIPO Copyright Treaty adopted on December 20, 1996, and/or similar international agreements. e. Exceptions and Limitations means fair use, fair dealing, and/or any other exception or limitation to Copyright and Similar Rights that applies to Your use of the Licensed Material. f. Licensed Material means the artistic or literary work, database, or other material to which the Licensor applied this Public License. g. Licensed Rights means the rights granted to You subject to the terms and conditions of this Public License, which are limited to all Copyright and Similar Rights that apply to Your use of the Licensed Material and that the Licensor has authority to license. h. Licensor means the individual(s) or entity(ies) granting rights under this Public License. i. NonCommercial means not primarily intended for or directed towards commercial advantage or monetary compensation. For purposes of this Public License, the exchange of the Licensed Material for other material subject to Copyright and Similar Rights by digital file-sharing or similar means is NonCommercial provided there is no payment of monetary compensation in connection with the exchange. j. Share means to provide material to the public by any means or process that requires permission under the Licensed Rights, such as reproduction, public display, public performance, distribution, dissemination, communication, or importation, and to make material available to the public including in ways that members of the public may access the material from a place and at a time individually chosen by them. k. Sui Generis Database Rights means rights other than copyright resulting from Directive 96/9/EC of the European Parliament and of the Council of 11 March 1996 on the legal protection of databases, as amended and/or succeeded, as well as other essentially equivalent rights anywhere in the world. l. You means the individual or entity exercising the Licensed Rights under this Public License. Your has a corresponding meaning. Section 2 -- Scope. a. License grant. 1. Subject to the terms and conditions of this Public License, the Licensor hereby grants You a worldwide, royalty-free, non-sublicensable, non-exclusive, irrevocable license to exercise the Licensed Rights in the Licensed Material to: a. reproduce and Share the Licensed Material, in whole or in part, for NonCommercial purposes only; and b. produce, reproduce, and Share Adapted Material for NonCommercial purposes only. 2. Exceptions and Limitations. For the avoidance of doubt, where Exceptions and Limitations apply to Your use, this Public License does not apply, and You do not need to comply with its terms and conditions. 3. Term. The term of this Public License is specified in Section 6(a). 4. Media and formats; technical modifications allowed. The Licensor authorizes You to exercise the Licensed Rights in all media and formats whether now known or hereafter created, and to make technical modifications necessary to do so. The Licensor waives and/or agrees not to assert any right or authority to forbid You from making technical modifications necessary to exercise the Licensed Rights, including technical modifications necessary to circumvent Effective Technological Measures. For purposes of this Public License, simply making modifications authorized by this Section 2(a) (4) never produces Adapted Material. 5. Downstream recipients. a. Offer from the Licensor -- Licensed Material. Every recipient of the Licensed Material automatically receives an offer from the Licensor to exercise the Licensed Rights under the terms and conditions of this Public License. b. No downstream restrictions. You may not offer or impose any additional or different terms or conditions on, or apply any Effective Technological Measures to, the Licensed Material if doing so restricts exercise of the Licensed Rights by any recipient of the Licensed Material. 6. No endorsement. Nothing in this Public License constitutes or may be construed as permission to assert or imply that You are, or that Your use of the Licensed Material is, connected with, or sponsored, endorsed, or granted official status by, the Licensor or others designated to receive attribution as provided in Section 3(a)(1)(A)(i). b. Other rights. 1. Moral rights, such as the right of integrity, are not licensed under this Public License, nor are publicity, privacy, and/or other similar personality rights; however, to the extent possible, the Licensor waives and/or agrees not to assert any such rights held by the Licensor to the limited extent necessary to allow You to exercise the Licensed Rights, but not otherwise. 2. Patent and trademark rights are not licensed under this Public License. 3. To the extent possible, the Licensor waives any right to collect royalties from You for the exercise of the Licensed Rights, whether directly or through a collecting society under any voluntary or waivable statutory or compulsory licensing scheme. In all other cases the Licensor expressly reserves any right to collect such royalties, including when the Licensed Material is used other than for NonCommercial purposes. Section 3 -- License Conditions. Your exercise of the Licensed Rights is expressly made subject to the following conditions. a. Attribution. 1. If You Share the Licensed Material (including in modified form), You must: a. retain the following if it is supplied by the Licensor with the Licensed Material: i. identification of the creator(s) of the Licensed Material and any others designated to receive attribution, in any reasonable manner requested by the Licensor (including by pseudonym if designated); ii. a copyright notice; iii. a notice that refers to this Public License; iv. a notice that refers to the disclaimer of warranties; v. a URI or hyperlink to the Licensed Material to the extent reasonably practicable; b. indicate if You modified the Licensed Material and retain an indication of any previous modifications; and c. indicate the Licensed Material is licensed under this Public License, and include the text of, or the URI or hyperlink to, this Public License. 2. You may satisfy the conditions in Section 3(a)(1) in any reasonable manner based on the medium, means, and context in which You Share the Licensed Material. For example, it may be reasonable to satisfy the conditions by providing a URI or hyperlink to a resource that includes the required information. 3. If requested by the Licensor, You must remove any of the information required by Section 3(a)(1)(A) to the extent reasonably practicable. 4. If You Share Adapted Material You produce, the Adapter's License You apply must not prevent recipients of the Adapted Material from complying with this Public License. Section 4 -- Sui Generis Database Rights. Where the Licensed Rights include Sui Generis Database Rights that apply to Your use of the Licensed Material: a. for the avoidance of doubt, Section 2(a)(1) grants You the right to extract, reuse, reproduce, and Share all or a substantial portion of the contents of the database for NonCommercial purposes only; b. if You include all or a substantial portion of the database contents in a database in which You have Sui Generis Database Rights, then the database in which You have Sui Generis Database Rights (but not its individual contents) is Adapted Material; and c. You must comply with the conditions in Section 3(a) if You Share all or a substantial portion of the contents of the database. For the avoidance of doubt, this Section 4 supplements and does not replace Your obligations under this Public License where the Licensed Rights include other Copyright and Similar Rights. Section 5 -- Disclaimer of Warranties and Limitation of Liability. a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS, IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION, WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS, ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU. b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION, NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT, INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES, COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR IN PART, THIS LIMITATION MAY NOT APPLY TO YOU. c. The disclaimer of warranties and limitation of liability provided above shall be interpreted in a manner that, to the extent possible, most closely approximates an absolute disclaimer and waiver of all liability. Section 6 -- Term and Termination. a. This Public License applies for the term of the Copyright and Similar Rights licensed here. However, if You fail to comply with this Public License, then Your rights under this Public License terminate automatically. b. Where Your right to use the Licensed Material has terminated under Section 6(a), it reinstates: 1. automatically as of the date the violation is cured, provided it is cured within 30 days of Your discovery of the violation; or 2. upon express reinstatement by the Licensor. For the avoidance of doubt, this Section 6(b) does not affect any right the Licensor may have to seek remedies for Your violations of this Public License. c. For the avoidance of doubt, the Licensor may also offer the Licensed Material under separate terms or conditions or stop distributing the Licensed Material at any time; however, doing so will not terminate this Public License. d. Sections 1, 5, 6, 7, and 8 survive termination of this Public License. Section 7 -- Other Terms and Conditions. a. The Licensor shall not be bound by any additional or different terms or conditions communicated by You unless expressly agreed. b. Any arrangements, understandings, or agreements regarding the Licensed Material not stated herein are separate from and independent of the terms and conditions of this Public License. Section 8 -- Interpretation. a. For the avoidance of doubt, this Public License does not, and shall not be interpreted to, reduce, limit, restrict, or impose conditions on any use of the Licensed Material that could lawfully be made without permission under this Public License. b. To the extent possible, if any provision of this Public License is deemed unenforceable, it shall be automatically reformed to the minimum extent necessary to make it enforceable. If the provision cannot be reformed, it shall be severed from this Public License without affecting the enforceability of the remaining terms and conditions. c. No term or condition of this Public License will be waived and no failure to comply consented to unless expressly agreed to by the Licensor. d. Nothing in this Public License constitutes or may be interpreted as a limitation upon, or waiver of, any privileges and immunities that apply to the Licensor or You, including from the legal processes of any jurisdiction or authority. ======================================================================= Creative Commons is not a party to its public licenses. Notwithstanding, Creative Commons may elect to apply one of its public licenses to material it publishes and in those instances will be considered the โ€œLicensor.โ€ The text of the Creative Commons public licenses is dedicated to the public domain under the CC0 Public Domain Dedication. Except for the limited purpose of indicating that material is shared under a Creative Commons public license or as otherwise permitted by the Creative Commons policies published at creativecommons.org/policies, Creative Commons does not authorize the use of the trademark "Creative Commons" or any other trademark or logo of Creative Commons without its prior written consent including, without limitation, in connection with any unauthorized modifications to any of its public licenses or any other arrangements, understandings, or agreements concerning use of licensed material. For the avoidance of doubt, this paragraph does not form part of the public licenses. Creative Commons may be contacted at creativecommons.org. Project-URL: Homepage, https://github.com/facebookresearch/seamless_interaction Project-URL: Bug Tracker, https://github.com/facebookresearch/seamless_interaction Classifier: Programming Language :: Python :: 3 Classifier: License :: OSI Approved :: CC BY-NC 4.0 Classifier: Operating System :: OS Independent Requires-Python: >=3.10 Description-Content-Type: text/markdown License-File: LICENSE Requires-Dist: boto3>=1.38.27 Requires-Dist: datasets>=3.6.0 Requires-Dist: fsspec>=2025.3.0 Requires-Dist: jupyterlab>=4.4.3 Requires-Dist: librosa>=0.11.0 Requires-Dist: llvmlite>=0.43.0 Requires-Dist: mypy>=1.16.1 Requires-Dist: numpy>=1.26.0 Requires-Dist: opencv-python>=4.7.0.72 Requires-Dist: pandas>=1.5.1 Requires-Dist: plotly>=6.1.2 Requires-Dist: pre-commit>=4.2.0 Requires-Dist: pylance>=0.29.0 Requires-Dist: pytest>=8.4.1 Requires-Dist: ruff>=0.12.0 Requires-Dist: s3fs>=0.4.2 Requires-Dist: streamlit>=1.46.0 Requires-Dist: webdataset>=1.0.2 Requires-Dist: wget>=3.2
# Seamless Interaction Dataset Seamless Interaction Dataset Banner **A large-scale multimodal dataset of 4,000+ hours of human interactions for AI research**
๐Ÿ–ผ๏ธ Blog ๐ŸŒ Website ๐ŸŽฎ Demo ๐Ÿค— HuggingFace ๐Ÿ“„ Paper
## Overview Human communication involves a complex interplay of verbal and nonverbal signals, essential for conveying meaning and achieving interpersonal goals. The **Seamless Interaction Dataset** is a large-scale collection of over 4,000 hours of face-to-face interaction footage from more than 4,000 participants in diverse contexts. This dataset enables the development of AI technologies that understand human interactions and communication, unlocking breakthroughs in: - ๐Ÿค– Virtual agents and embodied AI - ๐ŸŽญ Natural human-computer interaction - ๐Ÿ“ก Advanced telepresence experiences - ๐Ÿ“Š Multimodal content analysis tools - ๐ŸŽฌ Animation and synthetic content generation ## ๐Ÿš€ Quick Start ```bash git clone https://github.com/facebookresearch/seamless_interaction cd seamless_interaction pip install -e . streamlit run src/seamless_interaction/app/Welcome.py # if you use uv uv sync uv run streamlit run src/seamless_interaction/app/Welcome.py ``` Explore the dataset with our interactive browser: **Features:** - ๐Ÿ” **Hierarchical Navigation**: Browse by Label โ†’ Split โ†’ Batch โ†’ Interaction - ๐ŸŽฒ **Random Sampling**: Discover interactions with one-click random selection - ๐Ÿ“ฅ **Download Interface**: Download specific batches with size estimation and progress tracking - ๐ŸŽฌ **Video Viewer**: Side-by-side participant videos with synchronized playback - ๐Ÿ“Š **Data Analysis**: Overview statistics and distribution plots - ๐Ÿ“ **File Management**: Organize and preview audio, JSON, and NPZ files with expandable dropdowns ### Download Options We provide comprehensive download methods supporting all research scales and requirements: | **Scale** | **Size** | **Method** | **Use Case** | **Script** | **Sampling** | |-----------|----------|------------|--------------|------------|-------------| | ๐Ÿ” **Single Example** | ~100MB | S3 | Quick exploration, understanding data structure | [`download_s3.py`](./scripts/download_s3.py#L10) | Auto-sample from preferred vendors | | ๐Ÿ‘ฅ **Interaction Pair** | ~200MB | S3 | Study conversational dynamics between participants | [`download_s3.py`](./scripts/download_s3.py#L34) | Auto-detect conversation pairs | | ๐Ÿ“‚ **Sample Set** | ~1GB | S3/HF | Initial prototyping, algorithm development | [`download_s3.py`](./scripts/download_s3.py#L66), [`download_hf.py`](./scripts/download_hf.py#L10) | File selection or archive-based | | ๐ŸŽฏ **Session Groups** | ~400MB | S3 | Deep conversational context, session dynamics | [`download_s3.py`](./scripts/download_s3.py#L100) | Auto-sample rich sessions | | ๐Ÿ“ฆ **Single Batch** | ~50GB | HF | Substantial local development, full exploration | [`download_hf.py`](./scripts/download_hf.py#L24) | WebDataset tarball download | | ๐Ÿ—‚๏ธ **Multiple Batches** | ~150GB+ | HF | Training datasets, large-scale analysis | [`download_hf.py`](./scripts/download_hf.py#L38) | WebDataset tarball download | | ๐ŸŽฏ **Different Splits** | Variable | HF | Cross-validation (train/dev/test, improvised/naturalistic) | [`download_hf.py`](./scripts/download_hf.py#L55) | WebDataset tarball download | | ๐ŸŒ **Whole Dataset** | ~27TB | HF | Complete research dataset, production systems | [`download_hf.py`](./scripts/download_hf.py#L82) | WebDataset tarball download | #### ๐Ÿ” **S3 Download** - Fine-grained exploration Perfect for exploring individual interactions or specific file IDs. Downloads from S3 and automatically converts to consistent format (.wav, .mp4, .npz, .json). ```python # Initialize with configuration for cleaner setup from seamless_interaction.fs import SeamlessInteractionFS, DatasetConfig config = DatasetConfig( label="improvised", split="dev", preferred_vendors_only=True, local_dir=Path.home() / "datasets/seamless_interaction", # note: we will automatically create the directory if it doesn't exist ) fs = SeamlessInteractionFS(config=config) # Or use defaults: fs = SeamlessInteractionFS() file_ids = fs.sample_random_file_ids(num_samples=1) fs.gather_file_id_data_from_s3(file_ids[0]) # Or specify exact file: fs.gather_file_id_data_from_s3("V00_S0809_I00000582_P0947") # Files are organized as: # local_dir/improvised/train/0000/0005/V00_S0809_I00000582_P0947.wav # local_dir/improvised/train/0000/0005/V00_S0809_I00000582_P0947.mp4 # local_dir/improvised/train/0000/0005/V00_S0809_I00000582_P0947.json # local_dir/improvised/train/0000/0005/V00_S0809_I00000582_P0947.npz ``` For more details, please refer to the [S3 Download Example](./scripts/download_s3.py). #### ๐Ÿ“ฆ **HuggingFace Download** - Batch exploration Ideal for downloading self-contained batches (~50GB each) for local exploration. Each batch contains complete interaction pairs. ```python from seamless_interaction.fs import SeamlessInteractionFS, DatasetConfig # Initialize with configuration config = DatasetConfig(label="improvised", split="dev") fs = SeamlessInteractionFS(config=config) # Sample set (~1GB) - Quick exploration on laptops fs.download_batch_from_hf(batch_idx=0, archive_list=[0]) # Single batch (~50-100GB) - Substantial local development fs.download_batch_from_hf(batch_idx=0) ``` For more details, please refer to the [HuggingFace Download Example](./scripts/download_hf.py). ### Working with Downloaded Data ```python from seamless_interaction.fs import SeamlessInteractionFS import json import numpy as np import cv2 import librosa # Load interaction data def load_interaction_data(file_id): """Load all modalities for a given file ID.""" fs = SeamlessInteractionFS() paths = fs.get_path_list_for_file_id_local(file_id) print(paths) data = {} for path in paths: if path.endswith('.mp4'): data['video'] = cv2.VideoCapture(path) elif path.endswith('.wav'): data['audio'], data['sample_rate'] = librosa.load(path, sr=48_000) elif path.endswith('.json'): with open(path) as f: data['json'] = json.load(f) elif path.endswith('.npz'): data['npz'] = np.load(path) return data fs.download_archive_from_hf( archive=0, label="improvised", split="test", batch=0, local_dir=None, extract=True, ) # Example usage interaction = load_interaction_data("V01_S0223_I00000127_P1505") print(f"Available feature keys: {list(interaction['npz'].keys())}") print(f"Right hand pose data shape: {interaction['npz']['smplh:right_hand_pose'].shape}") ``` ### Basic Data Loading (HF + WebDataset) ```python from datasets import load_dataset # configure label = "improvised" split = "dev" batch_idx = 0 archive_list = [0, 1] base_url = ( f"https://huggingface.co/datasets/facebook/" f"seamless-interaction/resolve/main/{label}/{split}/" "{batch_idx:04d}/{archive_idx:04d}.tar" ) urls = [base_url.format(batch_idx=batch_idx, archive_idx=archive_idx) for archive_idx in archive_list] dataset = load_dataset( "webdataset", data_files={split: urls}, split=split, streaming=True ) for item in dataset: break isinstance(item["mp4"], bytes) # True item["npz"].keys() # dict_keys(['boxes_and_keypoints:box', 'boxes_and_keypoints:is_valid_box', 'boxes_and_keypoints:keypoints', 'movement:EmotionArousalToken', 'movement:EmotionValenceToken', 'movement:FAUToken', 'movement:FAUValue', 'movement:alignment_head_rotation', 'movement:alignment_translation', 'movement:emotion_arousal', 'movement:emotion_scores', 'movement:emotion_valence', 'movement:expression', 'movement:frame_latent', 'movement:gaze_encodings', 'movement:head_encodings', 'movement:hypernet_features', 'movement:is_valid', 'smplh:body_pose', 'smplh:global_orient', 'smplh:is_valid', 'smplh:left_hand_pose', 'smplh:right_hand_pose', 'smplh:translation']) item["json"].keys() # dict_keys(['id', 'metadata:transcript', 'metadata:vad']) item["wav"].keys() # dict_keys(['path', 'array', 'sampling_rate']) ``` Check out the [dataloader_webdataset.py](./scripts/dataloader_webdataset.py) script for more details. ## ๐Ÿ” Description The `seamless_interaction` repository is split into several main components: ### ๐Ÿ“Š Dataset The repository provides comprehensive tools for downloading, processing, and utilizing the Seamless Interaction dataset for research and development. The dataset includes: - **Raw and processed multimodal data**: Video, audio, transcripts, and annotations - **Precomputed features**: Motion capture, facial keypoints, voice activity detection - **Metadata**: Participant personality (BFI-2), interaction contexts, and relationships ### ๐Ÿ“‚ Repository Structure ``` seamless_interaction/ โ”œโ”€โ”€ assets/ # Static assets for documentation โ”‚ โ”œโ”€โ”€ banner.png โ”‚ โ””โ”€โ”€ filelist.csv # File list for the dataset โ”œโ”€โ”€ scripts/ # Example scripts for dataset usage โ”‚ โ”œโ”€โ”€ dataloader_webdataset.py โ”‚ โ”œโ”€โ”€ download_hf.py โ”‚ โ”œโ”€โ”€ download_s3.py โ”œโ”€โ”€ src/seamless_interaction/ # Main package source code โ”‚ โ”œโ”€โ”€ app/ # Data exploration application โ”‚ โ”œโ”€โ”€ fs.py # Filesystem interface for dataset access โ”‚ โ”œโ”€โ”€ utils.py # General utility functions โ”‚ โ”œโ”€โ”€ constants.py # Dataset constants and configuration โ”‚ โ””โ”€โ”€ __init__.py โ”œโ”€โ”€ LICENSE # CC-BY-NC 4.0 license โ””โ”€โ”€ pyproject.toml # Python package configuration ``` ## ๐Ÿ“ฆ Deep Dive into the Dataset ### Dataset Structure The Seamless Interaction Dataset is organized into two main categories/labels: - **Improvised**: Interactions primarily based on predefined scenarios with guided prompts with at least one professional actor. - **Naturalistic**: Prompted conversations that can be carried out by normal people. ``` seamless_interaction โ”œโ”€โ”€ improvised # Interactions with guided prompts โ”‚ โ”œโ”€โ”€ dev โ”‚ โ”‚ โ”œโ”€โ”€ 1P-IS/ # First-party internal state annotations โ”‚ โ”‚ โ”‚ โ””โ”€โ”€ V_S_I_P.json โ”‚ โ”‚ โ”œโ”€โ”€ 1P-R/ # First-party internal state rationale annotations โ”‚ โ”‚ โ”‚ โ””โ”€โ”€ V_S_I_P.json โ”‚ โ”‚ โ”œโ”€โ”€ 3P-IS/ # Third-party internal state annotations โ”‚ โ”‚ โ”‚ โ””โ”€โ”€ V_S_I_P.json โ”‚ โ”‚ โ”œโ”€โ”€ 3P-R/ # Third-party internal state rationale annotations โ”‚ โ”‚ โ”‚ โ””โ”€โ”€ V_S_I_P.json โ”‚ โ”‚ โ”œโ”€โ”€ 3P-V/ # Third-party visual annotation โ”‚ โ”‚ โ”‚ โ””โ”€โ”€ V_S_I_P.json โ”‚ โ”‚ โ”œโ”€โ”€ audio/ # Speaker-bleed denoised audio โ”‚ โ”‚ โ”‚ โ””โ”€โ”€ V_S_I_P.wav โ”‚ โ”‚ โ”œโ”€โ”€ boxes_and_keypoints/ โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ box/ # Bounding boxes for each participant โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ is_valid_box/ # Whether bounding boxes are valid โ”‚ โ”‚ โ”‚ โ””โ”€โ”€ keypoints/ # Detected facial/body keypoints โ”‚ โ”‚ โ”œโ”€โ”€ movement/ # Quantified Imitator movement features โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ emotion_arousal/ # Arousal measures โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ emotion_valence/ # Valence measures โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ emotion_scores/ # Emotion detection scores โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ expression/ # Facial expression parameters โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ FAUToken/ # Facial Action Unit tokens โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ FAUValue/ # Facial Action Unit values โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ gaze_encodings/ # Eye gaze direction encodings โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ head_encodings/ # Head position/rotation encodings โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ frame_latent/ # Per-frame latent representations โ”‚ โ”‚ โ”‚ โ””โ”€โ”€ is_valid/ # Validity flags for extracted features โ”‚ โ”‚ โ”œโ”€โ”€ smplh/ # SMPL-H body model parameters โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ body-pose/ # Body pose parameters โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ global_orient/ # Global orientation parameters โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ is_valid/ # Valid frames indicators โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ left_hand_pose/ # Left hand pose parameters โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ right_hand_pose/ # Right hand pose parameters โ”‚ โ”‚ โ”‚ โ””โ”€โ”€ translation/ # Global translation parameters โ”‚ โ”‚ โ”œโ”€โ”€ transcript/ # Time-aligned speech transcription โ”‚ โ”‚ โ”‚ โ””โ”€โ”€ V_S_I_P.jsonl โ”‚ โ”‚ โ”œโ”€โ”€ vad/ # Voice activity detection โ”‚ โ”‚ โ”‚ โ””โ”€โ”€ V_S_I_P.jsonl โ”‚ โ”‚ โ””โ”€โ”€ video/ # Raw HD video recordings โ”‚ โ”‚ โ””โ”€โ”€ V_S_I_P.mp4 โ”‚ โ”œโ”€โ”€ test/ # Test split with similar structure โ”‚ โ””โ”€โ”€ train/ # Training split with similar structure โ””โ”€โ”€ naturalistic/ # Spontaneous conversations โ”œโ”€โ”€ dev/ # Same structure as improvised/dev โ”œโ”€โ”€ test/ # Same structure as improvised/test โ””โ”€โ”€ train/ # Same structure as improvised/train ``` Each file is named according to a consistent convention: - `V`: Collection site/vendor identifier - `S`: Unique session identifier - `I`: Specific interaction within a session - `P`: Individual participant identifier ### Available Modalities and Features Each interaction in the dataset includes: | Modality | Description | File Format | Sample Rate | |----------|-------------|-------------|-------------| | ๐ŸŽฅ Video | High-definition face-to-face footage | MP4 (H.264) | 30/29.97 FPS, 1080p | | ๐ŸŽ™๏ธ Audio | Denoised audio with separate channels | WAV | 48kHz, 16-bit | | ๐Ÿ“ Transcript | Time-aligned speech transcription | JSONL | - | | ๐Ÿƒ SMPL-H | 3D body model parameters | NPY | 30 Hz | | ๐Ÿง  Imitator Movement Features | Comprehensive quantified imitator movement data | NPY | 30 Hz | | ๐Ÿ“Š Annotations | Human-annotated behavioral data | JSON | - | | ๐Ÿ”Š VAD | Voice activity detection | JSONL | 100 Hz | | ๐Ÿ“ฆ Keypoints | Face and body keypoints | NPY | 30 Hz | #### Annotation Types The dataset includes several types of human annotations for rich behavioral analysis: | Annotation | Hours | Total Annotations | Mean # Tokens | |------------|-------------|--------|--------| | 1P-IS (1st-party internal state annotations) | 1.1 | 751 | 5.8 | | 1P-R (1st-party internal state rationale annotations) | 1.1 | 751 | 10.2 | | 3P-IS (3rd-party internal state annotations) | 4.7 | 5132 | 5.2 | | 3P-R (3rd-party internal state rationale annotations) | 4.7 | 5132 | 11.3 | | 3P-V (3rd-party visual annotation) | 4.7 | 5132 | 14.6 | Please refer to the [technical report](https://ai.meta.com/research/publications/seamless-interaction-dyadic-audiovisual-motion-modeling-and-large-scale-dataset/) for a more detailed overview of annotations. #### Movement/Imitator Feature Types The movement directory contains rich behavioral features (output of the Imitator model): | Feature | Description | |---------|-------------| | `emotion_arousal` | Arousal intensity measurements | | `emotion_valence` | Valence (positive/negative) measurements | | `emotion_scores` | Detected emotion categorical scores | | `expression` | Parametric facial expression encodings | | `FAUToken`/`FAUValue` | Facial Action Unit tokens and intensity values | | `gaze_encodings` | Neural encodings of gaze direction | | `head_encodings` | Neural encodings of head position and rotation | | `frame_latent` | Per-frame latent representations | | `alignment_head_rotation` | Head rotation data for temporal alignment | | `alignment_translation` | Translation parameters for temporal alignment | | `EmotionArousalToken`/`EmotionValenceToken` | Discretized emotion tokens | | `hypernet_features` | Features from hypernetwork processing | ### Download Strategy Guide We provide two complementary download methods optimized for different research workflows: | Method | Use Case | Best For | Download Size | Parallelization | |--------|----------|----------|---------------|-----------------| | **S3 Direct** | Fine-grained exploration | Individual interactions, interaction pairs | Per file (~100MB) | โœ… Multiprocessing | | **HuggingFace Batches** | Batch processing | Local dataset exploration, model training | ~50-100GB per batch | โœ… Multiprocessing | #### When to Use S3 Download - **Qualitative analysis**: Examining specific interactions in detail - **Pair studies**: Analyzing conversational dynamics between participants - **Feature exploration**: Understanding data structure before large downloads - **Development**: Testing code with minimal data #### When to Use HuggingFace Download - **Model training**: Need substantial training data - **Batch processing**: Analyzing patterns across many interactions - **Local exploration**: Want self-contained dataset on laptop/workstation - **Reproducible research**: Ensure consistent data splits #### Performance Optimization ```python # Optimal settings for different systems config_default = DatasetConfig() # auto-detects system resources config_laptop = DatasetConfig(num_workers=4) # Laptop/small workstation config_workstation = DatasetConfig(num_workers=8) # High-end workstation config_server = DatasetConfig(num_workers=16) # Server/cluster node # Memory-efficient batch processing config = DatasetConfig(label="improvised", split="train") fs = SeamlessInteractionFS(config=config) for batch_idx in range(10): # Process in chunks fs.download_batch_from_hf(batch_idx=batch_idx) # Process batch here... # Delete batch to free space if needed ``` ### Dataset Versions The dataset is organized in self-contained batches for flexible exploration: | Split | Batches | Size per Batch | Total Size | Description | |-------|---------|----------------|------------|-------------| | **dev** | 5 | ~50GB | ~500GB | Development/validation set | | **test** | 5 | ~50GB | ~500TB | Hold-out test set | | **train** | 200+ | ~50GB | ~20TB+ | Full training data | #### Recommended Download Strategies ```python # Strategy 1: Quick Start (Laptop-friendly) config = DatasetConfig(label="improvised", split="dev") fs = SeamlessInteractionFS(config=config) fs.download_batch_from_hf(batch_idx=0, archive_list=[0, 1, 2]) # ~6GB # Strategy 2: Research Dataset (Workstation) config = DatasetConfig(label="improvised", split="dev") fs = SeamlessInteractionFS(config=config) fs.download_batch_from_hf(batch_idx=0) # Full dev set ~50-100GB config = DatasetConfig(label="naturalistic", split="dev") fs = SeamlessInteractionFS(config=config) fs.download_batch_from_hf(batch_idx=0) # Both interaction types # Strategy 3: Production Training (Server/Cluster) config = DatasetConfig(label="improvised", split="train") fs = SeamlessInteractionFS(config=config) for batch_idx in range(20): # First 20 training batches (~1TB) fs.download_batch_from_hf(batch_idx=batch_idx) ``` #### File Format Specifications Our data is stored in the following formats for optimal usability: | Format | Description | Usage | |--------|-------------|-------| | NPZ | NumPy array files | Efficient storage of numerical feature vectors, keypoints, and parameters | | JSONL | JSON Lines | Time-aligned annotations with one event per line (e.g., transcripts, VAD) | | JSON | JavaScript Object Notation | Structured metadata and annotations with timestamps | | MP4 | MPEG-4 Part 14 | High-quality compressed video with H.264 encoding | | WAV | Waveform Audio | Uncompressed audio for highest fidelity processing | ## ๐Ÿงช Research Applications The Seamless Interaction Dataset enables research across multiple domains: ### Embodied AI and Virtual Agents - Train agents that display natural gestures - Model turn-taking dynamics and interaction rhythms - Generate contextually appropriate responses to human behavior ### Multimodal Understanding - Analyze cross-modal correlations between speech, gesture, and expressions - Extract behavioral patterns from large-scale interaction data - Develop models to understand social dynamics ### Human-Computer Interaction - Design interfaces that respond to subtle human cues - Improve telepresence technologies with better behavioral modeling - Create more natural conversational agents ### Animation and Content Creation - Generate realistic human behaviors for animated characters - Synthesize conversational dynamics for virtual production - Create training data for digital human technologies ## โš ๏ธ Known Limitations and Noise in Metadata Given the scale and complexity involved in collecting the Seamless Interaction dataset, there are several known limitations that we will address in our ongoing work, with improvements planned for in future versions: ### Errors in Human-Based Time-Stamping The core unit of the dataset is interactions. An interaction defines the *active time* during which a participantโ€™s conversation and behavior can be linked to a pair of prompts. We have observed instances of misaligned time-stamps, including: - Annotated start/end times may be too early or too late. - Occasional misalignment between prompt text and spoken material. - Ordering of prompts that may contain off-by-one errors. Despite our efforts to automatically identify and correct these errors, approximately 10% of the interactions remain affected. ### Time Stamping "Noise" in Moments of Interest (MOI) While defining a MOI inherently involves some subjectivity, there are rare instances where: - The described behavior only represents a subset of the observed behavior. - The duration of the MOI does not fully capture the annotated behavior. ### Incorrect Assignment of Participant IDs In rare instances, we have observed: - Duplicate participant identifiers being assigned to different individuals. - The same individual being mapped to different identifiers. ### Unreleased "Meta Time" Currently, the dataset only contains *active time* segments - time in which two participants are actively responding to prompts. *Meta time* refers to the time between *active segments* in which participants are studying their new prompts, taking a break, etc. *Meta time* constitutes hundreds of hours in the raw collection and maybe be explored for future releases. ### Variation in Recording Site Consistency This multi-site project contains variation in: - Recording quality, including issues like speaker bleed and participants staying in frame. - Acting quality in *Improvised* segments. - The likelihood of time-stamping errors. All vendors met our technical requirements; however,there is noticeable variation in production quality across different sites. ## ๐Ÿค Contributing We welcome contributions from the research community! Here are some ways to contribute: - **Bug Reports & Feature Requests**: Open issues on GitHub - **Dataset Improvements**: Help enhance our preprocessing pipelines or annotations - **Model Contributions**: Submit your models to our benchmarks - **Documentation**: Improve our guides, tutorials, and API documentation - **Sample Code**: Share example applications built with the dataset Please see our [CONTRIBUTING.md](CONTRIBUTING.md) for detailed guidelines, code of conduct, and submission processes. ## ๐Ÿ“„ License & Data Usage Policy The Seamless Interaction Dataset is licensed under CC-BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International). This means you are free to: - **Share** โ€” copy and redistribute the material in any medium or format - **Adapt** โ€” remix, transform, and build upon the material Under the following terms: - **Attribution** โ€” You must give appropriate credit, provide a link to the license, and indicate if changes were made. - **NonCommercial** โ€” You may not use the material for commercial purposes without explicit permission. ## ๐Ÿ“‘ Citation If you use the Seamless Interaction Dataset in your research, please cite:
BibTeX ```bibtex @article{seamless_interaction, title={Seamless Interaction: Dyadic Audiovisual Motion Modeling and Large-Scale Dataset}, author={Vasu Agrawal and Akinniyi Akinyemi and Kathryn Alvero and Morteza Behrooz and Julia Buffalini and Fabio Maria Carlucci and Joy Chen and Junming Chen and Zhang Chen and Shiyang Cheng and Praveen Chowdary and Joe Chuang and Antony D'Avirro and Jon Daly and Ning Dong and Mark Duppenthaler and Cynthia Gao and Jeff Girard and Martin Gleize and Sahir Gomez and Hongyu Gong and Srivathsan Govindarajan and Brandon Han and Sen He and Denise Hernandez and Yordan Hristov and Rongjie Huang and Hirofumi Inaguma and Somya Jain and Raj Janardhan and Qingyao Jia and Christopher Klaiber and Dejan Kovachev and Moneish Kumar and Hang Li and Yilei Li and Pavel Litvin and Wei Liu and Guangyao Ma and Jing Ma and Martin Ma and Xutai Ma and Lucas Mantovani and Sagar Miglani and Sreyas Mohan and Louis-Philippe Morency and Evonne Ng and Kam-Woh Ng and Tu Anh Nguyen and Amia Oberai and Benjamin Peloquin and Juan Pino and Jovan Popovic and Omid Poursaeed and Fabian Prada and Alice Rakotoarison and Alexander Richard and Christophe Ropers and Safiyyah Saleem and Vasu Sharma and Alex Shcherbyna and Jia Shen and Jie Shen and Anastasis Stathopoulos and Anna Sun and Paden Tomasello and Tuan Tran and Arina Turkatenko and Bo Wan and Chao Wang and Jeff Wang and Mary Williamson and Carleigh Wood and Tao Xiang and Yilin Yang and Zhiyuan Yao and Chen Zhang and Jiemin Zhang and Xinyue Zhang and Jason Zheng and Pavlo Zhyzheria and Jan Zikes and Michael Zollhoefer }, url={https://ai.meta.com/research/publications/seamless-interaction-dyadic-audiovisual-motion-modeling-and-large-scale-dataset/}, year={2025} } ```
## ๐Ÿ™ Acknowledgments This project was made possible thanks to contributions from: - The thousands of participants who provided interaction data - Our dedicated annotation and QA team - Research collaborators from multiple institutions - FAIR (Fundamental AI Research) - The open-source community for valuable tools and libraries - Our data collection partners across multiple sites - Meta Reality Labs for supporting this research initiative