Less is More: Label-Guided Efficient Summarization of Procedural Videos

Shreya Rajpal, Michal Golovanevsky, and Carsten Eickhoff.

PRISM is a label-guided, training-free framework for procedural-video summarization. It processes fewer than 5% of video frames while retaining more than 84% semantic content and improves over baselines by up to 7.5%.