README.md

Narrations / Steps to Video Aligner (NSVA)

NSVA is a lightweight Transformer-based architecture, where the narration or procedural steps are used as queries, to iteratively attend the video features, and output the alignability or optimal temporal windows.

[project page] [Arxiv] [GitHub]

We provide pre-trained models for HTM-Align and HT-Step. You can use these two models for reproducing our results, following our [code].