HRDX dataset: Largest vector HD map for autonomous driving released

The HRDX dataset, introduced by Sahith Reddy Chada, Isht Dwivedi, and Nirav Savaliya, addresses the critical need for high-quality vector HD maps in autonomous driving. Existing public datasets are limited in scale and semantic richness, often lacking modalities like aerial imagery. HRDX is several times larger than prior datasets, offering about 40 hours (1,400 km) of minimally overlapping driving data. It was captured using six synchronized surround cameras, a 128-beam LiDAR, and centimeter-level RTK GNSS/IMU, with precisely aligned aerial orthoimagery. Annotations include 10 vector map classes and over 20 semantic and topological attributes, enabling a more comprehensive evaluation.

To assess performance on this richer ontology, the authors introduce the Composite Score (CS), which jointly evaluates geometric fidelity and attribute correctness. Benchmark experiments demonstrate that HRDX's scale significantly improves online vector-map construction. Furthermore, using aligned aerial imagery during training or inference boosts geometric map quality, and aerial-augmented teachers can transfer this benefit to camera-only students without extra sensor requirements at inference. The dataset supports reproducible research on large-scale HD-map learning, multimodal BEV fusion, and training-time privileged information. HRDX is publicly available with benchmarks.