Egocentric Vision for Accessibility AI

First-person video recorded by contributors with a range of accessibility needs as they navigate daily life, captured with rich scene and motion metadata. The collection exists so assistive vision and navigation models are trained on the people who actually rely on them, not a proxy population.

Every clip is collected from paid contributors with informed, accessible consent, scene-level provenance attached, and non-consenting bystanders blurred during QA before delivery.

Highlights

• First-person footage from contributors with diverse accessibility needs
• Rich multimodal metadata: scene, motion, and on-frame context
• Real daily-life navigation and task scenarios
• Informed, accessible consent built into the collection process
• Bystander blurring and PII review before delivery

Scenario coverage

Daily-life navigation and task scenarios across indoor and outdoor settings. Coverage extends to specific needs, environments, or tasks on request.

Capture and format

First-person video at 1920×1080 or higher and 30+ fps with synchronised motion metadata. Clips retain the full task and include scene-level context.

Annotations

Per-clip scene, location, and motion metadata as standard, plus optional object, hazard, and sub-step labels on request.

Provenance

• Paid contributors with informed, accessible consent
• Bystander faces blurred during QA before delivery
• No third-party copyrighted content on screen
• Per-clip audit trail and licensable usage rights

Use cases

• Assistive navigation and obstacle-detection models
• Scene description and object-finding for accessibility
• Inclusive evaluation of vision-language assistants
• Egocentric perception research with underrepresented users