license: other
license_name: spdo-data-use-agreement
license_link: https://huggingface.co/datasets/spdo-benchmark/spdo-eval/blob/main/DUA.md
pretty_name: 'SPDO: Same Place, Different Outcome'
task_categories:
- video-classification
- object-detection
language:
- en
tags:
- driving
- naturalistic-driving
- dashcam
- telematics
- hard-braking
- causal-attribution
- driver-monitoring
- benchmark
size_categories:
- n<1K
extra_gated_prompt: >-
Review copy of the SPDO evaluation set. Access requires agreement to the Data
Use Agreement (DUA.md).
extra_gated_fields:
Full name: text
Institution: text
Institutional email: text
Intended use (2-3 sentences): text
I agree to the Data Use Agreement and will not attempt to re-identify participants, vehicles or locations: checkbox
SPDO: Same Place, Different Outcome
A benchmark for attributing hard-braking events to road agents, situations and driver state from naturalistic dual-view driving video with synchronized telematics.
SPDO contains 866 accelerometer-verified hard-braking events recorded by 50 older drivers in their own vehicles over five years (2021–2025), each with a road-facing clip, a driver-facing clip, a synchronized telematics slice, and matched control passes: frames of the same driver traversing the same place without braking hard. A 237-event evaluation set carries object-level cause labels (bounding box on the causal object), a cause class, an external-vs-driver attribution and a driver-attention label, annotated by two annotators with κ = 0.80 agreement on cause. Full details are given in the accompanying paper.
Motivation
Telematics systems flag millions of hard brakes, but a flag records only that a driver braked hard, not why. Attribution of an event to a road agent (a lead vehicle, a vehicle cutting in, a pedestrian), a situation (a signal, a road feature) or the driver (an own maneuver, inattention) is what safety analysis, insurance and clinical driving assessment require. SPDO provides the data to evaluate this capability under naturalistic conditions, with event onsets defined by the vehicle's sensors rather than by annotators, and with repeat passes of the same place that separate the location from what happened there.
Composition
| Item | Value |
|---|---|
| Drivers (vehicles) | 50 (257 screened; 82 with hard-brake candidates; 50 with verified events and usable video) |
| Recording period | 2021–2025 (dates shifted per driver in the release) |
| Events | 866 accelerometer-confirmed hard brakes (≤ −0.3 g for ≥ 0.3 s) |
| Splits (by driver) | train 461 / val 102 / test 303 events; 19 / 13 / 18 drivers |
| Labeled evaluation events | 237 (val 80, test 157) from 30 drivers; 179 with a causal-object box |
| Events with same-place control frames | 789 (91%); 1,899 control passes, 93% same driver |
| Clip length | 15 s per view (−10 to +5 s around the GPS flag); onset time provided |
| Video | 1280×720 road and driver views; overlay bands masked; faces blurred; text and plates pixelated |
| Telematics | relative time, GPS speed, heading, accelerometer; no coordinates |
| Light | 782 day, 51 dawn/dusk, 33 night |
Cause classes in the evaluation set: signal/stop 89, lead vehicle 68, own maneuver 54, crossing vehicle 8, cut-in 7, road feature 5, vulnerable road user 4, none visible 2.
Files
events/<driver>/<id>_road.mp4, <id>_driver.mp4 event clips (ids: D07-e012)
telematics/<id>.csv t_rel_s, speed_kph, heading, accelerometer columns
controls/<id>/c<j>_t<offset>.jpg same-place frames of control pass j at the event's -4..+1 s positions
controls/controls.csv per pass: same_driver, days_apart, offsets, sync source
labels/labels_eval.csv cause, side, attention, box (normalized x1,y1,x2,y2 at onset), confidence, notes
labels/annotation_tasks.csv evaluation events and double-annotation flags
index/events.csv driver, split, shifted date, time of day, light, onset, peak g, speed, spot type
README.md, DUA.md, guideline.pdf
The braking onset is given in index/events.csv (accel_onset_rel_s, seconds relative to the GPS flag at clip time t = 10 s, i.e. clip time = 10 + accel_onset_rel_s).
Collection and processing
Each participant's vehicle was instrumented with a consumer dual-camera DVR and a telematics unit (GPS 1 Hz, accelerometer ≈25 Hz, OBD) with independent clocks. GPS-flagged hard-brake candidates were filtered (≥ 2 s, ≥ 20 km/h, ≤ 0.7 g) and confirmed against the accelerometer after per-trip axis calibration; 86% of filtered candidates are confirmed, and only 46% of raw GPS flags. Video was synchronized to the telematics through a per-week clock offset and an overlay-OCR fine synchronization validated to within ±1 s for 85% of verifiable labeled events. Control passes were found by GPS (within 25 m and 30° of heading, same driver first), re-verified brake-free with the accelerometer (5.5% of GPS-clean passes removed), and their frames extracted at the same positions as the event frames.
Annotation
Two annotators labeled all 237 evaluation events independently following the written guideline (guideline.pdf): cause class (counterfactual rule: the agent or situation without which the braking would not have occurred), a single tight box on the causal object at the onset frame, external-vs-driver side, and driver attention at onset. Agreement before adjudication: cause 0.86 (κ = 0.80); external-vs-driver κ = 0.68 where both annotators chose one of the two (0.55 with both counted as external); attention κ = 0.76; box center agreement 78%. Disagreements on cause (34 events) were adjudicated jointly; the released labels are the adjudicated set.
Intended use
Evaluation of hard-brake attribution: cause classification, causal-object localization, external-vs-driver attribution and driver attention; study of same-place contrast (event vs. control passes); driver-monitoring research on naturalistic data. Labels are provided for the val and test splits only; the train split is intended for self-supervised or weakly supervised learning using the sensor-defined onsets and the controls. Report test-split results; use val for tuning.
Known limitations
One region (South Florida, USA) and one population (older drivers); one camera system; 237 labeled events with a thin long tail (VRU, animal, road feature); driver attention unobservable in 39% of evaluation events because of sunglasses; control-frame timing uncertainty of about 1–2 s; traffic-light state not reliably readable at onset distance. Driver faces are blurred in the release.
Privacy and anonymization
Participants consented to the release of video with faces obscured and identifying information removed (IRB-approved). All driver-view frames are face-blurred; road-view frames have scene text and license plates pixelated and the camera's GPS/timestamp overlays masked; GPS coordinates are not released; participants and vehicles are identified only by anonymous driver ids; dates are shifted by a fixed random offset per driver; event ids carry no study identifiers. Users must not attempt to re-identify participants, vehicles or locations (see DUA.md).
Access
This review copy of the evaluation set is available on accepting the Data Use Agreement. The full release is gated: requests are reviewed by the study team and granted to researchers who agree to the Data Use Agreement for non-commercial research use. Code for the processing pipeline, the annotation tool, the baselines and the evaluation accompanies the paper as supplementary material.
Citation
@inproceedings{spdo2027,
title = {Same Place, Different Outcome: A Benchmark for Attributing Hard Braking in Naturalistic Dual-View Driving Video},
author = {[authors]},
booktitle = {[venue]},
year = {2027}
}
Maintenance and contact
Maintained by the study team. Contact details are withheld during review. Versioned releases are tagged in this repository; the label file carries a version number in its header.