European Dashcam Vehicle Classifier (v19)
Privacy-first dashcam vehicle recognition: detects and identifies make + model + generation of European cars, fully offline on Apple Silicon.
Model
- Architecture: ConvNeXt-Tiny (28M params), trained from scratch
- Classes: 646 European vehicle classes (make + model, incl. generations)
- Input: 224×224 RGB image of a vehicle crop
- Output: class logits → softmax probabilities
Honest evaluation
Measured on an intersection holdout (597 real dashcam crops that no model ever trained on) — this methodology was deliberately built after discovering that naive holdout comparisons were contaminated:
| Model | Top-1 | Top-5 |
|---|---|---|
| v19 (this model) | 57.2% | 79.7% |
| v20 | 56.5% | 79.4% |
| v21 | 54.7% | 77.9% |
| v22 | 53.7% | 79.9% |
Test-time augmentation (horizontal flip + scale averaging) adds ~+1pt.
Usage (PyTorch)
import torch
from torchvision import models
ckpt = torch.load("model.pt", map_location="cpu")
classes = ckpt["classes"]
model = models.convnext_tiny(weights=None)
model.classifier[2] = torch.nn.Linear(model.classifier[2].in_features, len(classes))
model.load_state_dict(ckpt["state_dict"])
model.eval()
# preprocess: resize to 224x224, normalize with ImageNet mean/std, predict
Training data
~21,300 real dashcam crops (extracted from public dashcam footage of European cities) + ~10,800 web images, 5x oversampled, from-scratch training. See the dataset card for provenance and licensing.
Privacy
The full system blurs license plates on stream + recordings by default (GDPR). This model operates on vehicle crops only.