JevSceneMiner
Mines timestamped driving scenes from vehicle logs by asking Jev lateral and longitudinal manoeuvre questions per moment.
Overview
JevSceneMiner turns recorded vehicle data into reviewable driving scenes. Input adapters convert motion, lane geometry and object interactions into a shared readable evidence schema, leaving missing evidence as unknown. For each moment, Jev answers lateral and longitudinal manoeuvre questions defined in plain language in an editable labels file, with probabilities per answer, and the results are merged into timestamped scenes such as turn right, then lane change, then keep lane, with speed phases inside each lateral scene.
A review tool shows synchronised camera, bird's-eye view, timeline, raw probabilities and ground-truth editing. The README reports one recorded Boston run on nuPlan data with Jev 1.13: 933 samples in 25.9 seconds with 8 workers, an estimated $0.32 total, with preprocessing and review local and only readable evidence sent to TypeSafe. Suggested uses are data curation, scenario-based training and validation, failure analysis and regression testing. A Medium write-up explains the approach.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What lateral manoeuvre is the vehicle doing at this moment?(inferred) | turn right · turn left · lane change · keep lane |
| Choice | What longitudinal phase is the vehicle in? | decelerating · cruising · accelerating |
How it works
- Adapters convert logs into shared readable evidence samples
- Jev answers lateral and longitudinal questions per sample in one request
- Moments are merged into timestamped scenes
- Reviewers inspect camera, BEV, timeline and probabilities and edit labels
Strengths
- Fast and cheap batch labelling per the recorded run
- No local inference GPU needed
- Plain-language editable label definitions
- Integrated evidence review tool
Caveats
- Figures from a single recorded run
- Evidence is sent to TypeSafe's API
- Accuracy depends on adapter quality and label wording
At a glance
- Type
- Project
- Status
- beta
- Language
- Python
- Runtime
- pipeline with review UI
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Mine driving scenes from logs with Jev: scene text in, scenarios with probabilities and timestamps out.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide?
The lateral and longitudinal manoeuvre at each moment of a driving log, with probabilities.
Does it run the model locally?
No. Preprocessing and review are local; readable evidence goes to TypeSafe's API.
What data was it shown on?
Real nuPlan footage from Boston in the demo.
Related projects
Pipeline that ingests Google Maps restaurant reviews and classifies sentiment, five rating pillars and flags with Jev or an LLM.
ServiceprototypeTypeScriptPython library with a TypeScript port that turns Jev questions into functions over lists and DataFrames, with caching and escalation.
LibrarybetaPython, TypeScriptFrench library ranking ADEME emission factors for an activity label, splitting code rules from Jev's semantic match.
ProjectprototypeJavaScriptResearch pipeline that scores NYC 311 complaint descriptions with Jev and maps severity and safety signals on a city grid.
ServiceresearchPythondbt package whose macros run Jev Choice, Noul and Score over SQL values in DuckDB and ClickHouse models.
ProjectbetaSQL (Jinja)DuckDB extension that asks Jev about every row in SQL and returns answers as ENUM, DOUBLE or STRUCT columns.
PluginbetaC++