Each row represents one target person in one video frame. The row contains the complete frame, the target person's bounding box, and the corresponding interaction-readiness label.

This public repository contains the processed interaction-readiness annotations and does not redistribute the original source images.

To reconstruct the complete dataset, users must obtain the corresponding source datasets directly from their original providers and comply with their respective licenses, access requirements, and data-use agreements.

The image_id, source_dataset, video_id, and frame_id fields can be used to associate each annotation with a locally obtained source frame.

An image-included version exists under https://huggingface.co/datasets/marilynchahine1/interaction-readiness-dataset which can be accessed after providing proof of granted access to SSUP-HRI, and proof of agreement to the terms and conditions of use of AVIDAR and JRDB.

Dataset configurations

The repository contains three separate configurations:

Configuration	Source dataset	File
`AVIDAR`	AVIDAR	`AVIDAR.parquet`
`JRDB`	JRDB	`JRDB.parquet`
`SSUP-HRI`	SSUP-HRI	`SSUP-HRI.parquet`

Each configuration is exposed as a separate dataset in the Hugging Face Dataset Viewer.

Dataset structure

Each row contains the following fields:

Column	Type	Description
`image`	Image	The complete video frame containing the target person.
`bbox`	Sequence of four float values	Target-person bounding box in `[x, y, width, height]` format.
`label`	String	Interaction-readiness annotation assigned to the target person.
`source_dataset`	String	Name of the original source dataset.
`video_id`	String	Identifier of the source video.
`person_id`	String	Identifier of the target person within the video.
`frame_id`	Integer	Frame identifier within the source video.

Example row:

{
    "image": <PIL.Image.Image>,
    "bbox": [245.0, 110.0, 102.0, 280.0],
    "label": "interaction_ready",
    "source_dataset": "AVIDAR",
    "video_id": "AVIDAR_001",
    "person_id": "1",
    "frame_id": 46,
}

Labels

The dataset uses the following interaction-readiness labels:

not_interaction_ready
interaction_ready
interaction_ongoing
interaction_done

Loading the dataset

Install the Hugging Face Datasets library:

pip install datasets

Load AVIDAR

from datasets import load_dataset

avidar = load_dataset(
    "marilynchahine/interaction-readiness",
    "AVIDAR",
    split="train",
)

Load JRDB

from datasets import load_dataset

jrdb = load_dataset(
    "marilynchahine/interaction-readiness",
    "JRDB",
    split="train",
)

Load SSUP-HRI

from datasets import load_dataset

ssup_hri = load_dataset(
    "marilynchahine/interaction-readiness",
    "SSUP-HRI",
    split="train",
)

Combining configurations

The configurations can be combined locally after loading:

from datasets import load_dataset, concatenate_datasets

avidar = load_dataset(
    "marilynchahine/interaction-readiness",
    "AVIDAR",
    split="train",
)

jrdb = load_dataset(
    "marilynchahine/interaction-readiness",
    "JRDB",
    split="train",
)

ssup_hri = load_dataset(
    "marilynchahine/interaction-readiness",
    "SSUP-HRI",
    split="train",
)

combined = concatenate_datasets(
    [
        avidar,
        jrdb,
        ssup_hri,
    ]
)

The source_dataset column can be used to identify the origin of each row after concatenation.

Accessing a sample

sample = avidar[0]

image = sample["image"]
bbox = sample["bbox"]
label = sample["label"]

print(type(image))
print(bbox)
print(label)

image.show()

The image field is decoded as a PIL image when accessed through the Hugging Face Datasets library.

Bounding-box format

Bounding boxes use the following format:

[x, y, width, height]

where:

x is the horizontal coordinate of the upper-left corner;
y is the vertical coordinate of the upper-left corner;
width is the width of the bounding box;
height is the height of the bounding box.

Coordinates are expressed in image pixels.

Dataset creation

The original annotations were converted into a unified person-per-frame representation.

For each target person and frame, the dataset stores:

the full video frame;
the target person's bounding box;
the person's interaction-readiness label;
source, video, person, and frame identifiers.

The image data is embedded in the Parquet files, so the original local frame paths are not required when loading the dataset.

Intended uses

The dataset is intended for research on:

interaction-readiness classification;
person-specific segmentation;
joint person localization and readiness classification;
human-robot interaction;
person-conditioned visual understanding;
multimodal and vision-language model evaluation;
zero-shot and few-shot interaction understanding.

Limitations

The annotations originate from multiple datasets with different environments, camera viewpoints, recording conditions, and interaction scenarios.

Potential users should account for:

class imbalance;
differences in annotation conventions between source datasets;
repeated frames when multiple people appear in the same image;
variation in image resolution;
domain differences between the source datasets.

The person_id values should only be interpreted within their corresponding video_id, unless otherwise documented.

Citation, licenses and source datasets

The Interaction Readiness dataset is only available for non-commercial use. Citation information for this processed dataset will be added following publication.

Please also cite the original datasets when using the corresponding configurations:

This repository contains data derived from AVIDAR, JRDB, and SSUP-HRI. The original images were combined with person-level annotations to create a person-per-frame interaction-readiness dataset.

Users are responsible for complying with the licenses, terms of use, and citation requirements of each original source dataset.

AVIDAR

The AVIDAR configuration in this repository was generated from the audio-visual sequences provided through: https://team.inria.fr/perception/avdiar/

@article{gebru2018audio, TITLE={Audio-Visual Speaker Diarization Based on Spatiotemporal Bayesian Fusion}, AUTHOR={Gebru, Israel D. and Ba, Sil{`e}ye and Li, Xiaofei and Horaud, Radu}, JOURNAL={IEEE Transactions on Pattern Analysis and Machine Intelligence}, VOLUME = {40}, NUMBER = {5}, PAGES = {1086-1099} YEAR = {2018}, DOI={10.1109/TPAMI.2017.2648793}, }

JRDB

The JRDB configuration in this repository was generated from the official JRDB Train Images and Test Images releases provided through: https://jrdb.erc.monash.edu/#downloads

@article{martin2021jrdb, title={JRDB: A Dataset and Benchmark of Egocentric Robot Visual Perception of Humans in Built Environments}, author={Martin-Martin, Roberto and Patel, Mihir and Rezatofighi, Hamid and Shenoi, Abhijeet and Gwak, JunYoung and Frankel, Eric and Sadeghian, Amir and Savarese, Silvio}, journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, volume={45}, number={6}, pages={6748--6765}, year={2023}, doi={10.1109/TPAMI.2021.3070543} }

License: Creative Commons Attribution-NonCommercial-ShareAlike 3.0

SSUP-HRI

Access to the SSUP-HRI videos was obtained through direct request and Data Usage Agreement (DUA) compliance. Dataset information is available at https://github.com/FAR-Lab/SSUP-HRI

It is necessary to be granted request by Frank (fb266@cornell.edu) for SSUP-HRI use in order to be granted access to this Interaction Readiness dataset. Please provide proof of granted access from Frank and the SSUP-HRI DUA agreement when requesting access to this dataset.

@inproceedings{bu2024ssuphri, author = {Fanjun Bu and Wendy Ju}, title = {SSUP-HRI: Social Signaling in Urban Public Human-Robot Interaction Dataset}, booktitle = {SS4HRI: Workshop on Social Signal Modelling}, year = {2024}, month = {March}, address = {Boulder, Colorado, USA}, publisher = {ACM} }

Contact

For questions about the processed annotations or dataset structure, contact the repository owner through Hugging Face.

Downloads last month: 10

Total file size:

89.1 GB