Research map / Detection and segmentation / Open-set and few-shot detection
eTraM: Event-based Traffic Monitoring Dataset
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces eTraM, a fully event-based traffic monitoring dataset from a static roadside Prophesee EVK4 camera (10 hr, ~2M bounding boxes, 8 classes), and benchmarks RVT, RED and YOLOv8 on it.
Problem
Event cameras remain largely unexplored for static traffic monitoring, and no fully event-based dataset from a static perspective existed for traffic participant detection under varied lighting and weather.
Approach
A first-of-its-kind static, fully event-based traffic dataset collected across intersections, roadways and local streets in day, twilight, night, and rainy conditions, with baselines and generalization tests on nighttime and unseen scenes.
Key results
- RVT achieves 0.309 overall AP50 and RED 0.303 on eTraM, versus 0.134 PED... YOLOv8 overall 0.197 across classes (vehicle/pedestrian/micro-mobility averaged per table)
- Adding 45 min of nighttime data to 2 hr daytime training raises RVT vehicle AP50 on nighttime test from 0.566 to 0.761 and RED vehicle from 0.374 to 0.673
- On an unseen intersection (held out), RVT vehicle AP50 is 0.628, higher than held-in 0.449
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use recurrent event-based detectors RVT and RED instead of frame-based YOLOv8 | eTraM traffic participant detection · eTraM test, Overall | AP50: RVT 0.539; RED 0.491; YOLOv8 0.178 | Helped |
| Add nighttime data to daytime training data | RVT · eTraM nighttime test, VH | AP50: 0.566 -> 0.761 | Helped |
| Add nighttime data to daytime training data | RVT · eTraM nighttime test, PED | AP50: 0.166 -> 0.254 | Helped |
| Add nighttime data to daytime training data | RED · eTraM nighttime test, VH | AP50: 0.374 -> 0.673 | Helped |
| Add nighttime data to daytime training data | RED · eTraM nighttime test, PED | AP50: 0.354 -> 0.422 | Helped |
| Evaluate on a held-out, unseen intersection rather than a held-in intersection | RVT · eTraM held-in vs held-out test, VH and PED | AP50: VH: 0.449 -> 0.628; PED: 0.316 -> 0.529 | Mixed |
| Evaluate on a held-out, unseen intersection rather than a held-in intersection | RED · eTraM held-in vs held-out test, VH and PED | AP50: VH: 0.556 -> 0.572; PED: 0.521 -> 0.509 | Mixed |
Surprising findings
- Event-based models trained only on daytime data do not generalize well to nighttime, despite event cameras being invariant to absolute illuminance.
- Held-out intersection test scores exceeded held-in scores for RVT on vehicles.
Limitations
- Data collected at a single university campus area in Tempe, Arizona, limiting geographic diversity.
- Nighttime data is noisy from reflections and light sources, and detection is weaker at intersections at night.
- Micro-mobility class performance varies widely due to few instances.
Experimental setup
Datasets: eTraM
Benchmarks: eTraM
Metrics: AP50
Baselines: YOLOv8
Builds on: RVT RED YOLOv8
Closest papers by meaning
- Traffic Scene Parsing through the TSP6K Dataset (CVPR 2024 · similarity 0.84)
- EventPillars: Pillar-based Efficient Representations for Event Data (AAAI 2025 · similarity 0.83)
- TraffiDent: A Dataset for Understanding the Interplay Between Traffic Dynamics and Incidents (NeurIPS 2025 · similarity 0.83)
- TUMTraf VideoQA: Dataset and Benchmark for Unified Spatio-Temporal Video Understanding in Traffic Scenes (ICML 2025 · similarity 0.83)
- Ev-3DOD: Pushing the Temporal Boundaries of 3D Object Detection with Event Cameras (CVPR 2025 · similarity 0.82)
- On-Road Object Importance Estimation: A New Dataset and A Model with Multi-Fold Top-Down Guidance (NeurIPS 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.