Reproduce “Let's Go Shopping (LGS) -- Web-Scale Image-Text Dataset for Visual Concept Understanding” 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
Yatong Bai, Utsav Garg, Apaar Shanker, Haoming Zhang, Samyak Parajuli, Erhan Bas, Isidora Filipovic, Amelia N. Chu and 5 more
The Let’s Go Shopping (LGS) dataset offers 15 million high-quality image-text pairs meticulously sourced from e-commerce websites, providing a public, web-scale resource for visual concept understanding. This dataset quantifies a significant domain shift between general natural images and e-commerce product images, demonstrating improved performance for tasks like self-supervised visual feature learning, attribute-rich image captioning, and e-commerce-styled text-to-image generation.
Key insights
- Commercial shopping websites offer a uniquely clean, informative, and fluent source for image-text data, overcoming limitations of general web-scraping methods.
- A significant and quantifiable distributional shift exists between general natural images and e-commerce product images, necessitating domain-specific datasets for robust AI performance.
- High-quality, domain-specific datasets like LGS can significantly improve self-supervised visual feature learning, image captioning, and text-to-image generation in relevant application areas.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.