Meitu Introduces a Multi‑Label Short‑Video Classification Dataset for the 2018 AI Challenger

On August 29, 2018, Meitu, Innovation Works, Sogou, and Meituan Dianping jointly launched the 2018 AI Challenger Global AI Competition. AI Challenger is currently the largest domestic platform for scientific datasets and the largest non‑commercial competition platform in China, emphasizing the integration of cutting‑edge research with industry practice and serving as a global AI talent community.

Building on the first edition, the competition invited more enterprises, universities, and government agencies, introducing more than a dozen new datasets across various fields and over a dozen contests with both research and industrial significance. This year Meitu participated as a co‑organizer and created a new track called “Real‑Time Short‑Video Classification,” inviting AI experts worldwide to showcase their skills.

Short‑video content has attracted extensive attention and rapid development in recent years. Meitu, as one of the earliest domestic companies to explore short‑video, launched the knowledge‑oriented short‑video community app Meipai in 2014. The platform now hosts rich content resources, covering over 300 knowledge domains and more than 30,000 short‑video creators.

Through years of operation, Meipai has accumulated massive short‑video data, which has been leveraged to strengthen Meitu’s video semantic understanding capabilities. Multi‑category, multi‑level, and multi‑dimensional video tagging technology has been successfully applied to content recommendation, intelligent operation, user profiling, and other scenarios.

For the 2018 AI Challenger, Meitu contributed the industry’s first multi‑label short‑video classification dataset. The dataset contains 200,000 short videos spanning 63 popular categories such as dance, fitness, and singing. It adopts a multi‑label taxonomy, with tags describing video subjects, scenes, actions, and other dimensions, covering almost every element shown in the videos.

Compared with traditional video datasets, this dataset has distinctive characteristics: the videos are primarily captured with mobile phones and are mostly vertical; many videos incorporate short‑video effects and richer editing operations; and the content is heavily centered on user‑generated selfie videos. These features reflect a user‑oriented content production trend.

Meitu states that it will continue to explore the short‑video field, using AI to tackle real‑world problems, further optimizing video understanding and analysis technologies while refining content quality. Meipai’s knowledge‑oriented short videos will evolve from broad categories to finer‑grained niches, creating new growth opportunities for the content ecosystem.

Note: This article is reproduced from Beijing Youth Daily.