CUSTOMVIDEOFIRST-PERSON
First-Person Activity Clips
Wearable-style first-person activity clips with task labels, step metadata, camera perspective, hands-visible review, environment fields, and optional narration for embodied AI evaluation.
CUSTOMVIDEODEXTEROUS TASKS
Hand and Object Interaction Clips
Controlled hand-object interaction collection format for robotic manipulation, dexterous hand modeling, affordance learning, and action understanding.
Hand and Object Interaction Clips are built for teams evaluating how models understand object affordances, hand motion, contact events, and manipulation sequences. The format captures short controlled clips where contributors perform defined actions on approved objects.
Structured collection
Hand and Object Interaction Clips use defined object-action tasks rather than open video capture.
Metadata depth
The format supports alignment between video timestamps, action labels, contact markers, object-state changes, and start/end…
Delivery-ready package
Includes video quality, object visibility, action match reviews, plus dataset manifest, CSV catalog, consent status metadata, QA…
Request sample access
Share your target object categories, action taxonomy, camera setup, annotation needs, and evaluation goals. We’ll reply with a technical scope for the collection.
Key highlights
- Controlled object manipulation clips with action verbs and object categories.
- Camera angle, surface type, lighting, and interaction setup metadata.
- Optional contact-event markers and object-state labels for manipulation analysis.
Dataset contents
Video capture
Video clips · Optional start/end action boundaries · Object visibility status fields · Object category labels · Object sub-type metadata
Steps, transcripts, and labels
Action verb labels · Hand dominance metadata · Interaction type metadata · Camera angle metadata · Surface type metadata
Capture and task metadata
Lighting metadata · Duration metadata · Optional contact-event markers · Optional object-state labels
Technical specifications
Use cases
- Robotic manipulation evaluation
- Dexterous hand modeling
- Affordance learning
- Action recognition
- Object state modeling
- Vision-language grounding
- Embodied model evaluation
- Contact event analysis
Dataset workflow
1Collection
Hand and Object Interaction Clips use defined object-action tasks rather than open video capture. Contributors record approved manipulation actions under controlled instructions for camera angle, object type, surface setup, and action sequence.
2Review
Video quality review · Object visibility review · Action match review · Contact marker review · Metadata review · Consent status review · Delivery approval
3Delivery
MP4 video clips · Object metadata CSV · Action metadata CSV · Annotation files where scoped · Dataset manifest JSON · CSV file catalog · QA summary · Consent status metadata · Delivery notes
Quality assurance
Related dataset formats
CUSTOMVIDEOINSTRUCTIONAL
Procedural Demonstration Video Sets
Step-structured procedural videos with task labels, timestamped step markers, narration alignment, completion status, and metadata for instruction-following and imitation learning.
CUSTOMVIDEOENVIRONMENT DIVERSITY
Multi-Environment Activity Clips
Activity video sets captured across varied real-world environments, with task labels, environment metadata, camera position fields, lighting conditions, duration, and QA status for activity recognition and embodied AI evaluation.
CUSTOMMULTIMODALPAIRED DATA
Multimodal Paired Capture Sets
Paired multimodal datasets linking video, audio, screen, text, or image-derived records through shared file IDs, cross-modal manifests, metadata schemas, and QA fields.
CUSTOMIMAGEPRODUCT RECOGNITION
Product and Packaging Image Sets
Product and packaging image sets with multi-angle capture, lighting metadata, package-side labels, category fields, resolution checks, QA status, and optional bounding-box or OCR annotations.
CUSTOMIMAGEOBJECT RECOGNITION
Everyday Object Photo Sets
Everyday object image sets with object IDs, multi-view capture, material/category metadata, angle labels, resolution checks, QA fields, and optional bounding boxes for perception models.