MA-EgoQA: 266 Hours of Multi-Agent Egocentric Video Question Answering

MA-EgoQA features 266 hours of multi-agent egocentric video and 1.7k question-answering pairs, developed by Kangsan Kim in 2026. The collection captures six individuals living together for seven days, providing a benchmark for simultaneous long-horizon video understanding.

Use Cases

• Multi-agent video question answering using the 1.7k question-answer pairs
• Long-horizon activity recognition across 266 hours of footage
• Social interaction analysis between six participants in a shared space

Strengths

• 266 hours of video data
• Synchronized multi-agent perspectives from 6 participants
• 1.7k annotated question-answering pairs

Limitations

• Small participant pool of 6 individuals
• Single-environment geographic bias restricted to an indoor living space
• Low label density relative to the total video duration

Provenance

Source

Kangsan Kim et al. (Arxiv: 2603.09827)

Collection Method

Egocentric camera recording and manual QA annotation

Time Range

7 days

Freshness

Last updated March 2026