BDD100K is a large-scale driving video dataset containing 100,000 videos and 10 tasks, designed to evaluate image recognition algorithms for autonomous driving. The dataset includes diverse geographic, environmental, and weather conditions, making it suitable for training models that are less likely to be surprised by new conditions. It provides a benchmark for heterogeneous multitask learning, enabling the study of how to solve various tasks together. The dataset includes tasks such as image tagging, lane detection, drivable area segmentation, road object detection, semantic segmentation, instance segmentation, multi-object detection tracking, multi-object segmentation tracking, domain adaptation, and imitation learning. The dataset is collected from over 50,000 rides, covering cities like New York and San Francisco, and includes GPS/IMU recordings to preserve driving trajectories. The dataset is split into training (70K), validation (10K), and testing (20K) sets. The dataset supports a wide range of tasks, including object detection, lane marking, drivable area detection, semantic instance segmentation, multiple object tracking, and segmentation. The dataset is diverse in terms of weather, scene types, and time of day, and includes annotations for various tasks. The dataset is used to evaluate existing algorithms and study the effects of multitask learning in homogeneous, cascaded, and heterogeneous settings. The results show that special training strategies are needed for existing models to perform such heterogeneous tasks. The dataset provides a comprehensive set of annotations for various tasks, enabling the study of heterogeneous multitask learning. The dataset is also used to evaluate the performance of models on different tasks and to study the challenges of designing a single model to support multiple tasks. The dataset is a valuable resource for future research in autonomous driving and computer vision.BDD100K is a large-scale driving video dataset containing 100,000 videos and 10 tasks, designed to evaluate image recognition algorithms for autonomous driving. The dataset includes diverse geographic, environmental, and weather conditions, making it suitable for training models that are less likely to be surprised by new conditions. It provides a benchmark for heterogeneous multitask learning, enabling the study of how to solve various tasks together. The dataset includes tasks such as image tagging, lane detection, drivable area segmentation, road object detection, semantic segmentation, instance segmentation, multi-object detection tracking, multi-object segmentation tracking, domain adaptation, and imitation learning. The dataset is collected from over 50,000 rides, covering cities like New York and San Francisco, and includes GPS/IMU recordings to preserve driving trajectories. The dataset is split into training (70K), validation (10K), and testing (20K) sets. The dataset supports a wide range of tasks, including object detection, lane marking, drivable area detection, semantic instance segmentation, multiple object tracking, and segmentation. The dataset is diverse in terms of weather, scene types, and time of day, and includes annotations for various tasks. The dataset is used to evaluate existing algorithms and study the effects of multitask learning in homogeneous, cascaded, and heterogeneous settings. The results show that special training strategies are needed for existing models to perform such heterogeneous tasks. The dataset provides a comprehensive set of annotations for various tasks, enabling the study of heterogeneous multitask learning. The dataset is also used to evaluate the performance of models on different tasks and to study the challenges of designing a single model to support multiple tasks. The dataset is a valuable resource for future research in autonomous driving and computer vision.