Reproduce “PREGO: online mistake detection in PRocedural EGOcentric videos” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Alessandro Flaborea, Guido Maria D'Amely di Melendugno, Leonardo Plini, Luca Scofano, Edoardo De Matteis, Antonino Furnari, Giovanni Maria Farinella, Fabio Galasso
PREGO introduces an online, one-class classification framework for detecting open-set procedural mistakes in egocentric videos. It integrates an online action recognition module with a Large Language Model for symbolic next-step anticipation, identifying errors when observed actions diverge from expected sequences, achieving an F1 score of 41.8% on the Assembly101-O dataset.
Key insights
- Abstract symbolic representations for actions, combined with in-context learning by Large Language Models, are effective for long-term procedural reasoning and next-step anticipation in an open-set, zero-shot manner.
- The accuracy of the online step recognition component is a primary bottleneck for overall mistake detection performance, indicating that improvements in real-time video analysis could significantly enhance the system.
- Structured prompt engineering and clear context delivery to the LLM are crucial for improving the precision and recall of symbolic action anticipation and subsequent mistake detection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.