First-Person Household Video Dataset Collection Case Study

Client

A global urban mobility technology company

Objective

The objective of this project was to record day-to-day household activities from a first-person point of view, using head-mounted cameras, to build a dataset for training AI models to understand real-world domestic tasks.

Methodology

All recordings were captured using issued head-mounted equipment only, ensuring a consistent point of view across every participant — no handheld or alternate-angle capture was permitted. Recording and upload were both routed through a purpose-built mobile application (Android-only), with fixed technical specifications: 0.5x zoom lens, 1080p resolution or higher at 30 FPS, landscape orientation.

Participants worked through a defined task list of common household activities, including:

• Collecting and sorting laundry

• Folding and putting away laundry

• Hanging clothes

• Ironing clothes

• Loading/unloading a dishwasher

• Loading/unloading a laundry machine

• Organising kitchen cabinets

• Tidying a room

• Washing dishes in the sink

• Organising a table

Each activity followed detailed, task-specific recording guidance shared separately with participants, along with reference training videos in their preferred language. Devices were switched to airplane mode before recording to avoid interruptions, and participants performed a quick self-QA replay before uploading. Recommended clip length was 5–15 minutes per video.

Upload Infrastructure

A strong Wi-Fi connection was recommended for automatic, efficient uploads; time spent uploading over mobile data (where Wi-Fi was unavailable) was excluded from billable effective hours. Devices required 10–15 GB of free storage daily, and videos were auto-deleted from the app once successfully uploaded.

Privacy and Quality Controls

• No personally identifiable information (PII) was permitted in any recording.

• Only the participant's own hands, up to the wrist, could appear in frame — no other people, hands, or reflections (including on TV/screens) were allowed.

• Content variety was enforced — for example, already-ironed clothing could not be reused unless it had naturally re-wrinkled.

• All submissions went through a formal QC process, with only QC-approved footage eligible for payment and feedback shared within roughly 3 days.

Key Learnings

• Standardising on issued head-mount equipment and app-only capture removed the variability that ad hoc phone recording would have introduced.

• The private, in-home nature of the recordings made clear participant training on privacy rules (no PII, no bystanders, hands-only up to the wrist) essential.

• A defined task list gave structure and comparability across participants, while still allowing for variety within each task category.

Need first-person video data for training text-to-image or editing models? Market Xcel can run app-based, privacy-compliant egocentric video collection across a PAN-India field network — reach out to scope your requirements.