Model
ViT Vehicle Classifier
Estimate vehicle make and support plate / speed attributes for tracked vehicles.
beta
Vehicle Attributes
VehiclesANPR
Overview
A vision transformer estimates vehicle make from tracked-vehicle crops, confidence-gated and voted across frames to stay stable. It runs alongside licence-plate estimation (ANPR) and an ego-motion-compensated speed estimate, so each vehicle carries subtype, make, plate and a rough km/h read.
| Role | Vehicle Attributes |
|---|---|
| Expected latency | ~8–15 ms / crop (GPU) |
| Hardware requirements | NVIDIA GPU recommended |
| Memory requirements | ~1–2 GB VRAM |
Inputs & Outputs
Inputs
- Vehicle crop
- Track id
- Frames for voting
Outputs
- Estimated make (confidence-gated)
- Plate string (ANPR)
- Speed estimate (km/h)
Advantages & Limitations
Advantages
- Adds rich vehicle attributes
- Voting stabilises noisy single-frame reads
- Complements plate + speed
Limitations
- Make limited to trained classes
- Plates need adequate resolution / angle
- Speed uncalibrated, comparable, not exact
Example outputs
ViT Vehicle Classifier
A tracked car labelled with make + plate + km/h
ViT Vehicle Classifier
A voted make that stabilises over several frames
Configuration
Keys in config/default.yaml that govern this model.
| Key | Default |
|---|---|
vehicle.make.live_interval | 4.0 |
vehicle.anpr.enabled | true |
Benchmarks
Placeholder
Benchmark figures below are illustrative and awaiting a standardised harness. Treat them as placeholders.
| Benchmark | Dataset | Metric | Value | Hardware |
|---|---|---|---|---|
| Make top-1 | internal | accuracy | placeholder | - |