Mining Multi-Modality Spatio-Temporal Cues for Video Important Person Identification

Researchers introduced a new video important person identification task to rank key people in video scenes.

They released Temporal-VIP, a rationale-annotated dataset with 9,249 labeled video segments.

VIP-Net combines social and temporal cues to improve person ranking and explain why someone is important.

The model reached 67.3% accuracy and produced stronger rationale generation for video understanding tasks.