Wreck-7K: 7,000-Hour Real-World Human–Object Interaction Video

Wreck-7K: 7,000-Hour Real-World Human–Object Interaction Video Archive

Data Product Features

• Third-Person Video: Exocentric recordings showing participants, objects, tools, and the surrounding environment.
• Natural Human Behavior: Unscripted activity recorded during real customer sessions rather than staged laboratory demonstrations.
• Human–Object Interaction: Repeated examples of grasping, carrying, positioning, striking, sweeping, shoveling, sorting, and collecting objects.
• Tool Use: Participants interact with a variety of handheld tools and physical objects.
• Impact and Breakage Events: Video includes object impacts, deformation, fragmentation, breakage, and changes in object condition.
• Object-State Transitions: Objects and scenes change from intact or organized states to damaged, fragmented, displaced, or collected states.
• Scene Transformation: The environment changes throughout each session as objects are moved, broken, scattered, and removed.
• Multi-Person Activity: Some sessions include multiple participants interacting within the same environment.
• Variable Occlusion: Participants, objects, tools, and debris may partially or fully occlude one another during activity.
• Long-Form Temporal Context: Continuous video captures action sequences and scene evolution rather than only isolated clips.
• Real-World Variation: The archive includes variation in movement style, participant height, clothing, object type, tool choice, action speed, and interaction intensity.
• Commercial Licensing: Qualified footage is available for commercial machine-learning training, evaluation, and model development.

Distribution

• Primary Format: Video files in their original or standardized delivery format
• Video Codec: H.264
• Resolution: 1920X1080
• Frame Rate: 30 FPS
• Audio: Removed
• Manifest Format: CSV and/or JSON
• Delivery Method: Secure cloud transfer, encrypted storage device, or another mutually agreed delivery method
• Folder Structure: Organized by session, camera, date, or asset identifier
• Annotations: Raw video is the standard deliverable. Additional segmentation, tagging, transcription, object detection, face blurring, audio removal, or metadata enrichment can be quoted separately.
• Data Volume: Approximately 7,000 hours of third-person video
• Number of Sessions: Thousands of individual customer sessions
• Number of Video Assets: Dependent on the final licensed subset and delivery configuration
• Manifest Records: One record per delivered video asset
• Estimated Storage Size: 12TB
• Ongoing Availability: Additional footage may be added through continued capture operations

Usage

• Video Foundation Models: Large-scale pretraining and evaluation using real-world human activity and physical interactions.
• Vision-Language Models: Training systems to connect video content with descriptions of actions, objects, events, and scene changes.
• Action Recognition: Identification and classification of carrying, striking, sweeping, shoveling, sorting, cleanup, and other physical actions.
• Temporal Action Segmentation: Detecting when actions begin, change, and end within long-form video.
• Human–Object Interaction Recognition: Modeling relationships among people, tools, objects, and the surrounding environment.
• Object-State Recognition: Identifying transitions such as intact to damaged, organized to scattered, or debris to collected.
• Physical-World Understanding: Learning how objects move, deform, break, scatter, and interact after physical contact.
• World-Model Development: Training or evaluating models that predict future visual states following human actions and object interactions.
• Scene Understanding: Analyzing changing layouts, clutter, irregular objects, occlusion, debris, and multi-person activity.
• Tool-Use Recognition: Detecting tools, tool selection, tool handling, and resulting interactions with target objects.
• Event Detection: Identifying impacts, breakage events, debris movement, cleanup actions, and other meaningful moments.
• Synthetic Video Evaluation: Comparing generated video, simulated physics, and synthetic activity against real-world examples.
• Robotics Visual Pretraining: Supporting affordance learning, activity understanding, visual representation learning, and human-to-robot transfer research.
• Safety and Behavior Analysis: Studying movement patterns, proximity, tool use, high-motion activity, and interactions in cluttered environments.

Coverage

• Geographic Coverage: 60% tourists in Las Vegas, Nevada, United States
• Capture Environment: Indoor operating physical entertainment facility
• Time Range: 2018 – Ongoing
• Session Coverage: 4000+ real customer sessions
• Participant Coverage: Thousands of participants across a wide range of body types, movement styles, clothing, group sizes, and activity patterns
• Age Coverage: Primarily adults. Exact demographic labels are not included unless separately documented and made available.
• Gender Coverage: Participants of multiple genders appear in the footage. Gender labels are not included as structured metadata.
• Environmental Coverage: Variable object configurations, debris levels, lighting conditions, participant positions, action intensity, and scene states within the same operating facility

License

• Evaluation access
• Curated hourly subsets
• Full non-exclusive commercial licensing
• Field-limited licensing
• Internal research and model-development licensing
• Ongoing data delivery
• Custom privacy processing
• Custom annotation and enrichment

AI Training Rights

• Use the Data Product to train, fine-tune, test, validate, and evaluate machine-learning models, including video models, computer-vision models, vision-language models, multimodal foundation models, and embodied-AI systems.
• Incorporate information learned from the Data Product into commercial models and commercialize resulting model outputs.
• Create permitted derivative works, including model weights, embeddings, representations, labels, and analytical outputs.
• Use the Data Product for internal research, model development, benchmarking, and evaluation.

• The raw Data Product may not be sold, redistributed, sublicensed, published, or shared outside the licensed organization except as expressly authorized.
• The Data Product may not be used to identify, contact, track, or profile individual participants.
• The licensee must comply with all applicable privacy, data-protection, intellectual-property, and machine-learning regulations.
• The licensee must maintain reasonable technical and organizational safeguards protecting the Data Product.
• Any permitted use remains subject to the final commercial license agreement.

Who Can Use It

• AI and Machine-Learning Companies: For training video, vision-language, multimodal, physical-world, and foundation models.
• Computer-Vision Teams: For action recognition, human–object interaction detection, object-state recognition, and scene understanding.
• Robotics Companies: For visual pretraining, affordance learning, task understanding, and human-demonstration research.
• Research Organizations: For academic or commercial research involving video understanding, physical interaction, human behavior, and temporal modeling.
• Simulation and Synthetic-Data Companies: For evaluating generated motion, object interactions, impact dynamics, breakage, and scene transformation.
• Video-Generation Companies: For training and evaluating models that generate or predict realistic human actions and physical-world events.
• Safety and Behavior Researchers: For analyzing tool use, movement, proximity, impacts, and activity in cluttered environments.
• Data Aggregators and Licensing Platforms: For authorized licensing and delivery to qualified enterprise customers, subject to contractual restrictions.

Data Dictionary

Additional Notes

• Curated subset creation
• Scene and action segmentation
• Object and activity tagging
• Automated object detection
• Tracking or segmentation masks
• Speech transcription
• Face blurring
• Audio removal
• Clip extraction
• Quality filtering
• Custom manifest development
• Ongoing directed data capture