Audi's A2D2 dataset offers 41K annotated frames for autonomous driving

Audi has open-sourced the A2D2 (Audi Autonomous Driving Dataset), a comprehensive multimodal dataset designed to accelerate research in machine learning, mobile robotics, and autonomous driving. The dataset comprises simultaneously recorded images and 3D point clouds from a sensor rig with six cameras and five LiDAR units, offering full 360-degree coverage. All sensor data is time-synchronized and mutually registered. Annotations include semantic segmentation and instance segmentation for 41,277 frames, plus 3D bounding boxes for objects in the front camera's field of view across 12,497 frames. The team also provides 392,556 sequential unannotated frames from loops driven in three southern German cities, useful for pretraining or self-supervised learning.

To comply with GDPR and protect anonymity, all faces and vehicle license plates have been blurred. The dataset is released under the CC BY-ND 4.0 license, which permits commercial use but prohibits derivative works. This licensing choice makes A2D2 a valuable resource for both industry and academia, as it can be directly used for training and evaluation without modification. The data and further information are available on the project website. With its high-quality annotations and multi-sensor setup, A2D2 fills a critical gap in the autonomous driving dataset landscape, following similar efforts like Waymo Open Dataset and nuScenes.