SODA10M is a semi-/self-supervised 2D benchmark dataset, which mainly contains 10 million diverse unlabeled road scene images and 20,000 images annotated with 6 representative object categories.

In order to maintain the rich diversity of data, unlabeled images were selected from 32 different cities, covering most areas of China. The images also contain a variety of different road scenes (cities, highways, urban and rural roads, parks), weather (sunny, cloudy, rainy, snowy), time periods (daytime, nighttime, early morning/dusk).