Benchmark RadarChehre
Chehre is a video dataset of 2,111 facial expressions prompted by 40 emojis, with annotations transferred to synthetic faces. It defines two tasks: dominant expression recognition and distributional expression recognition, evaluating models' ability to predict human-rated labels and capture response diversity.
- Released
- 2026-06-19
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Existing facial expression benchmarks rely on static images and basic categories, limiting evaluation of dynamic, diverse expressions. Chehre provides a controlled resource for measuring model performance on varied, distributional perception, with tasks that reveal gaps in current vision-language models.
Motivation
Facial expressions are nonverbal social signals used in human interaction, but facial expression recognition datasets often focus on static images, basic emotion categories, or single deterministic annotations.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.