A dataset for human–robot collaboration

Understanding intent.
From every angle.

Five camera views of two people working together. A closer look at the actions, hands and visual cues that make collaborative assembly possible.

QUB – Perception of Human Engagement in Assembly Operation

COLLECTION GEOMETRY01 — 05
Diagram of an assembly table, two people and five viewpoints. Illustration, not a recording.ULURLLLRAV
DYADIC ASSEMBLY · FIVE VIEWSIllustration
70participants in the study
36assembly subtasks
5camera viewpoints
50participants in the consented video release

01 / The dataset

Small actions.
Shared intentions.

QUB-PHEO captures collaborative assembly from an aerial camera and four side cameras, supporting research into how people anticipate and respond to one another.

The study combines multi-view recordings with fine-grained assembly subtask labels. The consented distribution is a subset of the original study; research access is governed by the dataset agreement.

01

BIAH

Assembly task family

02

Tower

Assembly task family

03

Stairway

Assembly task family

04

Bridge

Assembly task family

Dataset documentation and benchmarks ↗

02 / Interactive explorer

One moment. Five perspectives.

Inspect saved body, face, hand, object and gaze predictions alongside the video. Playback follows relative clip time.

Landmark reviewNo sample loaded
0.00 / 0.00 s
Saved landmark layers

Recordings require approved access. Already have the browser sample export? Open it locally to explore; its files stay on your computer.

Click a landmark to inspect its coordinates and stored confidence.

Same layout as the research viewer. UL and LL on the left, aerial in the centre, UR and LR on the right. Twenty representative examples cover four task families and twenty subtasks.

Interpreting the overlays. Saved predictions can be missing or inaccurate. Unknown confidence and actor identity remain unknown; synchronisation follows existing clip timing.

03 / Signals and provenance

Know what
you are looking at.

Landmark formats by camera view
ViewSaved signalsInterpretation
UL / UR133-point whole body, 478-point face, hands and objectsStored masks, model metadata and confidence where available.
LL / LR33-point body, 478-point face, 21-point handsLegacy pixel coordinates. Actor identity and confidence are unknown.
AerialHands, gaze and bounding boxesLegacy normalised coordinates. Generic detections carry no object class.

The browser samples are derivatives of matched videos and saved landmarks. Source hashes, frame counts, timestamp checks and transformation details accompany the export. They are demonstration examples, not an evaluation split.

Preprocessing and local visualiser code ↗

04 / Research access

Open to research.
Shared with care.

The full dataset stays on Google Drive. Access requires the official EULA and Seán McLoone’s approval. An application or signed agreement does not automatically grant access.

Read the official EULA ↗
  1. 01

    Read and complete the agreement

    Follow the official EULA’s electronic signing instructions. Use your full name and institutional email.

  2. 02

    Request research access

    Send the completed agreement to s.mcloone@qub.ac.uk, with your institution or group and intended use.

  3. 03

    Seán reviews your request

    Following approval, the team provides access to the appropriate consented data on Drive. Existing access arrangements remain in place.

05 / Research and citation

Built for the
next question.

DATASET PAPER · IEEE ACCESS · 2024

QUB-PHEO: A Visual-Based Dyadic Multi-View Dataset for Intention Inference in Collaborative Assembly

Samuel Adebayo, Seán McLoone and Joost C. Dessing
Volume 12, pages 157050–157066

View citation
@article{qub_pheo_2024,
  author = {Adebayo, Samuel and McLoone, Seán and Dessing, Joost C.},
  title = {QUB-PHEO: A Visual-Based Dyadic Multi-View Dataset for Intention Inference in Collaborative Assembly},
  journal = {IEEE Access},
  year = {2024},
  volume = {12},
  pages = {157050--157066},
  doi = {10.1109/ACCESS.2024.3485162}
}

Further research
Dyadic visual motion prediction baselines ↗