Driving-perception benchmarks often lack dense ground truth across both camera images and 3D space. This release addresses that gap with depth and semantic occupancy annotations for 1,253 clips, making it possible to evaluate how a system sees scene geometry as well as what occupies the vehicle’s surroundings.
What the Labels Enable
- Metric depth maps provide camera-view range targets, while voxel labels encode occupied regions, semantic classes, and velocity in the vehicle frame. This supports evaluation across complementary 2D and 3D perception tasks.
- The clips come from real-world vehicle fleets and build on NVIDIA’s PhysicalAI-Autonomous-Vehicles data. The annotations are generated automatically from temporally aggregated LiDAR and machine-generated labels, so they offer scalable targets but should not be mistaken for wholly human-verified ground truth.
- The benchmark is designed to validate perception models and autolabeling tools. Its fixed set of 1,253 clips makes comparisons reproducible, but the release is not a general-purpose driving dataset for unrestricted training or deployment.
Who It’s For—and the Trade-Off
A good fit for researchers benchmarking autonomous-vehicle depth, detection, or semantic occupancy systems against dense labels. Look elsewhere if you need commercial-use rights, unrestricted reuse, or a broad training corpus: the dataset is non-commercial and its license limits use to internal autonomous-vehicle and driver-assistance development using NVIDIA technology.
