Learning from PhotoShop Operation Videos: The PSOV Dataset

摘要

In this paper, we present the PhotoShop Operation Video (PSOV) dataset, a large-scale, densely annotated video database designed for the development of software intelligence. The PSOV dataset consists of 564 densely-annotated videos for Photoshop operations, covering more than 500 commonly used commands in the Photoshop software. Videos in this dataset are obtained from YouTube, manually watched and annotated precisely to seconds by experts. There are more than 74 h of videos with 29,204 labeled commands. To the best of our knowledge, the PSOV dataset is the first large-scale software operation video database with high-resolution frames and dense annotations. We believe that this dataset can help advance the development of intelligent software, and has extensive application aspects. In this paper, we describe the dataset construction procedure, data attributes, proposed tasks and their corresponding evaluation metrics. To demonstrate that the PSOV dataset has sufficient data and labeling for data-driven methods, we develop a deep learning based algorithm for the command classification task. We also carry out experiments and analysis with the proposed method to encourage better understanding and usage of the PSOV dataset.