Dataset Factory: A Toolchain For Generative Computer Vision Datasets

TL;DR

The paper introduces a 'dataset factory' toolchain that simplifies large-scale data management and processing for generative computer vision datasets, addressing challenges in data wrangling, sharing, and versioning.

Contribution

It presents a novel approach that separates data storage from metadata, enabling scalable data-centric operations for vision datasets.

Findings

Enables efficient handling of petabyte-scale datasets.

Supports robust dataset sharing and versioning.

Facilitates data filtering and processing at scale.

Abstract

Generative AI workflows heavily rely on data-centric tasks - such as filtering samples by annotation fields, vector distances, or scores produced by custom classifiers. At the same time, computer vision datasets are quickly approaching petabyte volumes, rendering data wrangling difficult. In addition, the iterative nature of data preparation necessitates robust dataset sharing and versioning mechanisms, both of which are hard to implement ad-hoc. To solve these challenges, we propose a "dataset factory" approach that separates the storage and processing of samples from metadata and enables data-centric operations at scale for machine learning teams and individual researchers.