DriveDNA: A Large-Scale Multimodal Naturalistic Driving Dataset and Benchmark for Driving Style Identification
DriveDNA tests whether models can recognize driver-specific behavior without confusing it for route, vehicle, or context.
The dataset covers 4,121 drives from 465 drivers, across 115 vehicle models and 975 hours of human-controlled driving. Its benchmark centers on few-shot driver re-identification, personalized behavior prediction, and condition-matched comparison. Learned representations beat classical descriptors on unseen drivers, while video-only models showed route leakage despite strong recognition results. HF Daily Papers' note
The dataset covers 4,121 drives from 465 drivers, across 115 vehicle models and 975 hours of human-controlled driving. Its benchmark centers on few-shot driver re-identification, personalized behavior prediction, and condition-matched comparison. Learned representations beat classical descriptors on unseen drivers, while video-only models showed route leakage despite strong recognition results. HF Daily Papers' note
score 5