COIN Behavior Recognition Dataset

Instructional videos can teach users how to complete corresponding tasks, but existing tutorial video datasets have great limitations in scale and diversity, making them difficult to apply to complex scenarios in real life. To this end, the COIN dataset was created for more comprehensive instructional video analysis.

The COIN dataset, also known as a large-scale dataset for COmprehensive INstructional video analysis, includes 11,827 videos collected from YouTube, covering 180 different tasks in 12 areas related to daily life (such as vehicles, gadgets, etc.).

The dataset contains 476 hours of video and 46,354 annotated segments. The average video length is 2.36 minutes, with 3.91 annotated step segments, and the average length of each segment is 14.91 seconds.

To download the dataset, please visit COIN official website.