21,458 video clips across 14 privacy-sensitive categories including faces, license plates, and digital screens. Annotations include bounding boxes for personally identifiable information (PII) captured in first-person video streams.
Use Cases
- Train a privacy-aware object detector using the bounding box coordinates and category labels.
- Benchmark automated de-identification algorithms on the 'face' and 'license_plate' classes.
- Analyze the frequency and duration of PII exposure in daily life using the temporal annotations.
- Develop privacy-preserving video summarization tools that mask the 'screen' and 'document' categories.
Strengths
- 21,458 annotated video segments derived from egocentric footage
- 14 privacy-sensitive categories including 'face', 'license_plate', and 'screen'
- 4.3 million bounding box annotations for sensitive objects
- Temporal consistency labels for tracking sensitive objects across consecutive frames