NARRATE
A Multimodal Real-World Australian Driving Dataset for Human-Centred Explanations in Automated Driving.
Dataset Showcase
Five annotated driving events with multimodal sensor data and natural language explanations.
Open dataset videoannotated driving events
drivers
sensor/data streams
DriveX Workshop archival track
Project Overview
NARRATE pairs on-road multi-sensor capture with participant questionnaires and span-level situational-awareness labels to support human-centred explanation research in automated driving.
The project includes camera, LiDAR, GNSS, IMU, telemetry, and natural language explanation data collected in real-world Australian driving contexts.
ECCV 2026 DriveX Workshop
At ECCV 2026 in Malmo, Sweden, I presented NARRATE with Tina Mehraban at the DriveX Workshop. After collecting, analysing, and shaping this dataset, it was meaningful to share the work with the research community and discuss how real drivers explain their actions.
These discussions connected directly to the project goal: supporting automated vehicles that communicate in a more cognitively human-centred way, grounded in how people naturally understand and describe driving situations.
Technical Contribution
- Designed and ran a real-world field study collecting 2,050 multimodal driving events from 35 drivers.
- Engineered Python ingestion pipelines for multi-sensor synchronization, cleaning, and participant-disjoint splitting.
- Benchmarked transformer backbones including RoBERTa, CLIP, T5, and BERT on perception and anticipation tasks.
Dataset Pipeline
