SODA10M

We introduce a new large-scale 2D dataset, named SODA10M, which contains 10M unlabeled images and 20k labeled images with 6 representative object categories. SODA10M is designed for promoting significant progress of self-supervised learning and domain adaptation in autonomous driving. It is the largest 2D autonomous driving dataset until now and will serve as a more challenging benchmark for the community.Self-supervised Learning for Next-generation Industry-level Autonomous driving refers to a variety of studies that attempt to refresh the solutions on challenging real-world perception tasks by learning from unlabeled or semi-supervised large-scale collected data to incrementally self-train powerful recognition models. Thanks to the rise of large-scale annotated data sets and the advance in computing hardware, various supervised learning methods have significantly improved the performance in many problems (e.g. 2D detection, instance segmentation and 3D Lidar Detection) in the field of self-driving. However, these supervised learning approaches are notorious "data hungry", especially in the current autonomous driving fields.The performance of self-driving perception systems highly relies on the annotation scale of labeled bounding boxes and IDs, which makes them not practical in many real-world industrial applications. The intuition is that a human driver can keep accumulating experiences from self-exploring the roads without any tutor’s guidance, instead current CV solutions are still baby-sitted with extensive annotation efforts on every new scenario.To facilitate an industry-level autonomous driving system in the future, the desired visual recognition model should be equipped with the ability of self-exploring, self-training and self-adapting across diverse new-appearing geographies, streets, cities, weather conditions, object labels, viewpoints or abnormal scenarios. To address this problem, many recent efforts in self-supervised learning, large-scale pretraining, weakly supervised learning and incremental/continual learning have been made to improve the perception systems to deviate from traditional paths of supervised learning for self-driving solutions.The aim of releasing this dataset is let the public to explore methods that utilizing both labeled data and unlabled data to achieve industry-level autonomous driving solutions. The benchmark paper has been released at Arxiv and it will be used to hold the ICCV2021 SSLAD chanllege