Datasets:
pretty_name: 'HarassGuard: Social VR Harassment Behavior Video Dataset'
license: other
license_name: harassguard-research-data-use-agreement
task_categories:
- video-classification
tags:
- video
- social-vr
- virtual-reality
- harassment-detection
- human-behavior
- vision-language-models
- controlled-access
- research-use-only
size_categories:
- 1K<n<10K
gated: true
extra_gated_heading: Request access to the HarassGuard Dataset
extra_gated_description: >-
Access is limited to approved researchers for non-commercial academic and
research use. Requests are reviewed manually by the dataset providers.
extra_gated_button_content: Agree and request access
extra_gated_prompt: |
The HarassGuard Dataset contains video data collected from human participants
under an IRB-approved user study. By requesting access, you confirm that the
information you provide is accurate and that, if access is granted, you will
comply with the HarassGuard Research Data Use Agreement (DUA). Access does not
permit redistribution or public release of the original video data.
extra_gated_fields:
Institution:
type: text
Department or Laboratory:
type: text
Position:
type: select
options:
- Faculty
- Researcher
- Postdoctoral Researcher
- Ph.D. Student
- M.S. Student
- Undergraduate Student
- Other
Principal Investigator or Faculty Supervisor:
type: text
Institutional Email Address:
type: text
Research Project Title:
type: text
Intended Research Use:
type: text
IRB or Research Ethics Status:
type: select
options:
- Approved
- Exempt
- Pending
- Not required by my institution
- Not yet determined
IRB or Ethics Protocol Number (if applicable):
type: text
I confirm that I am requesting access solely for non-commercial academic or research purposes:
type: checkbox
I have read and agree to comply with the HarassGuard Research Data Use Agreement:
type: checkbox
I will not redistribute, publicly upload, sublicense, or otherwise provide the original video data to third parties:
type: checkbox
I will not attempt to identify or re-identify any participant:
type: checkbox
I will restrict access to approved members of my research team and store the dataset securely:
type: checkbox
I understand that publication of original video clips, screenshots, or extracted frames requires prior written permission from the dataset provider:
type: checkbox
HarassGuard Dataset
Social VR Harassment Behavior Video Dataset
The HarassGuard Dataset is a vision-based video dataset developed for detecting physical harassment behaviors in Social Virtual Reality (Social VR) environments using visual information.
The dataset was collected through the user study conducted for the following paper:
HarassGuard: Detecting Harassment Behaviors in Social Virtual Reality with Vision-Language Models
Junhee Lee, Minseok Kim, Hwanjo Heo, Seungwon Woo, Jinwoo Kim
IEEE VR 2026
- IRB Approval: Kwangwoon University Institutional Review Board
(No. 7001546-20250630-HR(SB)-006-06) - Number of Video Clips: 1,116 (
.mp4) - Resolution / Frame Rate: 1280 × 720 / 30 fps (H.264)
- Data Collection Period: August 20–27, 2025
- Access: Controlled access for non-commercial academic and research use
1. Dataset Overview
The HarassGuard Dataset contains video recordings of avatar interactions in Social VR environments.
Harassment behaviors are organized into three major harassment categories, together with a Benign category representing normal interactions.
All behavior labels are defined based on interactions between avatars. Background activities, such as normal gameplay unrelated to harassment, are excluded from harassment labeling.
Behavior Categories
| Category | Subcategory | Definition |
|---|---|---|
| Aggressive Behavior | Punching | Striking another avatar with a fist |
| Slapping | Striking another avatar with an open hand | |
| Hitting with Object | Striking another avatar using an object | |
| Punching & Slapping | A combination of punching and slapping behaviors; present only in the Communication Room | |
| Personal Space Violation | Looming | Moving excessively close to another avatar's face |
| Following / Stalking | Continuously following another avatar | |
| Disruptive Behavior | Blocking | Physically obstructing another avatar's movement |
| Benign Behavior | — | Normal interactions such as conversation, rock-paper-scissors, walking, and other non-harassing activities |
2. Dataset Statistics
Number of Videos by Room and Category
| Room \ Category | Aggressive | Personal Space Violation | Disruptive | Benign | Total |
|---|---|---|---|---|---|
| Communication Room | 230 | 140 | 120 | 30 | 520 |
| Whack-a-Pig Room | 140 | 60 | — | 60 | 260 |
| Slingshot Room | 110 | 60 | — | 61 | 231 |
| Climbing Room | — | — | — | 105 | 105 |
| Total | 480 | 260 | 120 | 256 | 1,116 |
Aggressive Behavior Breakdown
- Communication Room
- Punching: 90
- Slapping: 60
- Hitting with Object: 60
- Punching & Slapping: 20
- Whack-a-Pig Room
- Hitting with Object: 70
- Punching: 40
- Slapping: 30
- Slingshot Room
- Punching: 80
- Slapping: 30
Personal Space Violation Breakdown
- Communication Room
- Looming: 80
- Following / Stalking: 60
- Whack-a-Pig Room
- Looming: 60
- Slingshot Room
- Looming: 60
3. Participant Demographics
The user study was conducted under approval from the Kwangwoon University Institutional Review Board (IRB).
A total of 14 participants took part in the study.
| Characteristic | Description |
|---|---|
| Number of Participants | 14 |
| Gender | 9 male, 5 female |
| Age | 18–34 years |
| Prior VR Experience | 11 experienced, 3 inexperienced |
| Most Familiar Platform | VRChat: 12; No prior platform experience: 2 |
| Prior Social VR Experience | 4 participants |
| Prior Experience Witnessing Harassment | 6 participants |
| Compensation | Approximately USD 10 per participant |
VR Usage Among Participants with Prior VR Experience
- Less than 1 hour/day: 7 participants (Beginner)
- 1–2 hours/day: 3 participants (Intermediate)
- 2–3 hours/day: 1 participant (Advanced)
4. Study Design
Session Organization
- Participants were organized into groups of 2–4 people per session.
- Group composition could vary during the experiment.
Experimental Roles
The study employed a within-subject design.
Each participant performed both the harasser and victim roles at least once during the experiment.
Duration
Each participant spent approximately 1 hour completing the study, including preparation and VR experience.
Pre-experiment Training
Participants received:
- approximately 5 minutes of training on the VR headset and controllers; and
- approximately 15 minutes of free exploration across the four experimental rooms.
5. Recording Environment
VR Environment
- Engine: Unity 2022.3.22f1
- Base Environment: Unity VR Multiplayer Template
- The environment was configured to support the VRChat development setting used in the study.
Avatar Tracking
Full-body avatars were represented using:
- upper-body tracking; and
- Inverse Kinematics (IK) to reproduce lower-body movements.
Camera Configuration
The recording environment consisted of:
- four third-person omnidirectional cameras (
3PoV_1–3PoV_4); and - one first-person camera (
1PoV).
All video streams were recorded directly on the server.
| Viewpoint | Number of Clips |
|---|---|
First-person (1PoV) |
224 |
Third-person (3PoV) |
892 |
| Total | 1,116 |
6. Dataset Structure
Dataset/
├── Communication_Room/ (520 clips)
│ ├── Aggressive/
│ │ ├── Punching/ (90)
│ │ ├── Slapping/ (60)
│ │ ├── Hitting_With_Object/ (60)
│ │ └── Punching&Slapping/ (20)
│ ├── Personal_Space_Violation/
│ │ ├── Looming/ (80)
│ │ └── Following/ (60)
│ ├── Disruptive_Behavior/
│ │ └── Blocking/ (120)
│ └── Benign/ (30)
│
├── Whackapig_Room/ (260 clips)
│ ├── Aggressive/
│ │ ├── Hitting_With_Object/ (70)
│ │ ├── Punching/ (40)
│ │ └── Slapping/ (30)
│ ├── Personal_Space_Violation/
│ │ └── Looming/ (60)
│ └── Benign/ (60)
│
├── Slingshot_Room/ (231 clips)
│ ├── Aggressive/
│ │ ├── Punching/ (80)
│ │ └── Slapping/ (30)
│ ├── Personal_Space_Violation/
│ │ └── Looming/ (60)
│ └── Benign/ (61)
│
└── Climbing_Room/ (105 clips)
└── Benign/ (105)
The Climbing Room contains only benign behaviors.
7. Video Specifications
| Property | Specification |
|---|---|
| Container | MP4 |
| Codec | H.264 |
| Resolution | 1280 × 720 |
| Frame Rate | 30 fps |
| Clip Duration | Variable; approximately 20 seconds or longer |
8. Controlled Access
Because the HarassGuard Dataset contains video data collected from human participants as part of an IRB-approved user study, the original video files are not available through unrestricted public access.
The dataset is distributed through the Hugging Face Hub as a gated dataset with manual review.
Access is intended for researchers affiliated with recognized universities, research institutes, or other bona fide research organizations who wish to use the dataset for non-commercial academic and research purposes.
Each access request is reviewed individually by the dataset providers. Approval is not automatic.
Access Request Information
Hugging Face automatically provides the requester's account username and email address to the dataset providers. The gated access form additionally requests:
- institution;
- department or laboratory;
- position;
- principal investigator or faculty supervisor;
- institutional email address;
- research project title;
- intended research use; and
- IRB / research ethics information, if applicable.
Applicants must also explicitly agree to the key requirements of the HarassGuard Research Data Use Agreement (DUA).
Use of an institutional email address is strongly recommended. The dataset providers may request additional information to verify institutional affiliation or the proposed research purpose.
9. Terms of Use
Access to the original video dataset is governed by the:
HarassGuard Research Data Use Agreement (DUA)
See DATA_USE_AGREEMENT.md for the complete terms.
In summary:
- The dataset may be used only for non-commercial academic and research purposes.
- Machine-learning and AI model training and evaluation are permitted.
- Normal research preprocessing, including frame extraction, resizing, cropping, annotation, and feature extraction, is permitted.
- Redistribution of the original videos is prohibited.
- Public uploading of the original videos or substantial portions of the dataset is prohibited.
- Researchers must not attempt to identify or re-identify study participants.
- Dataset access must be restricted to approved research-team members.
- The dataset must be stored using reasonable security measures.
- Publication of original video clips, screenshots, extracted frames, or other visual excerpts requires prior written permission from the dataset provider.
- Research publications using the dataset must appropriately cite the HarassGuard paper.
The above is only a summary. The complete terms in DATA_USE_AGREEMENT.md govern use of the original video dataset.
10. Derived Data and Research Outputs
Researchers may create and publish research outputs derived from the dataset, including:
- trained model parameters;
- aggregate statistics;
- experimental results;
- plots and tables;
- annotations;
- non-reconstructable embeddings or features; and
- algorithms and source code developed using the dataset.
Derived data that contains substantial portions of the original videos, permits practical reconstruction of the original videos, or may compromise participant privacy must not be publicly redistributed without prior written permission.
11. Research Ethics and Privacy
The HarassGuard Dataset originates from an IRB-approved human-participant study.
Researchers receiving the dataset are responsible for determining whether their proposed secondary use requires additional approval, exemption, or review by their own Institutional Review Board (IRB), Research Ethics Committee, or equivalent authority.
Researchers must not attempt to identify or re-identify any participant.
The dataset must not be used for surveillance, profiling, harassment, discrimination, or other purposes unrelated to the approved research.
Researchers are responsible for complying with all applicable institutional policies, ethical requirements, and laws governing their use of the dataset.
12. Data Security and Third-Party Services
Researchers must store the original video data using reasonable technical and organizational safeguards.
Original videos should not be transferred to third-party services, including external cloud-based AI or vision-language-model APIs, unless such transfer is permitted by:
- the HarassGuard Research Data Use Agreement;
- the researcher's applicable IRB / ethics approval or exemption;
- the researcher's institutional policies; and
- the third-party service's data-processing, retention, and privacy terms.
Researchers remain responsible for ensuring that any permitted third-party processing does not result in unauthorized retention, disclosure, redistribution, or use of the original video data.
13. Citation
If you use the HarassGuard Dataset in your research, please cite:
Junhee Lee, Minseok Kim, Hwanjo Heo, Seungwon Woo, and Jinwoo Kim.
HarassGuard: Detecting Harassment Behaviors in Social Virtual Reality with Vision-Language Models.
IEEE Transactions on Visualization and Computer Graphics, 2026.
BibTeX
@article{lee2026harassguard,
title={HarassGuard: Detecting harassment behaviors in social virtual reality with vision-language models},
author={Lee, Junhee and Kim, Minseok and Heo, Hwanjo and Woo, Seungwon and Kim, Jinwoo},
journal={IEEE Transactions on Visualization and Computer Graphics},
year={2026},
publisher={IEEE}
}
14. License and Data Use Agreement
The original video files are not released under an open-content license such as CC BY or CC BY-NC.
They are provided under the:
HarassGuard Research Data Use Agreement
Controlled Access — Non-commercial Academic and Research Use Only
Receiving access does not grant permission to redistribute, publicly upload, sublicense, or otherwise provide the original videos to third parties.
The public dataset card, label descriptions, aggregate statistics, and documentation may be referenced and reused with appropriate attribution, subject to any separately stated terms.
See DATA_USE_AGREEMENT.md for the complete conditions governing access and use.
15. Disclaimer
The dataset is provided "as is" without warranties of any kind, express or implied.
The dataset providers and their affiliated institutions make no representations regarding the accuracy, completeness, fitness for a particular purpose, or suitability of the dataset.
Researchers are responsible for ensuring that their use of the dataset complies with the HarassGuard Research Data Use Agreement and all applicable ethical, institutional, and legal requirements.
16. Contact
For questions regarding the dataset or requests requiring clarification beyond the Hugging Face gated-access form, please contact the HarassGuard Dataset authors.
Dataset Contact:
[Hwanjo Heo/hwanjo@etri.re.kr]
Paper:
HarassGuard: Detecting Harassment Behaviors in Social Virtual Reality with Vision-Language Models
IEEE VR 2026
Repository Files
README.md
DATA_USE_AGREEMENT.md
README.md— Dataset description, statistics, access policy, and Hugging Face gated-access formDATA_USE_AGREEMENT.md— Complete terms governing access and use of the original video dataset