Also detect and crop the vehicle with a pretrained YOLO model, then classify the crop. Full-image stays the official prediction. Falls back gracefully if the detector is unavailable.
Datasets: Stanford Cars · CompCars
Model: EfficientNet-B3 (fine-tuned, 300 px, Stanford Cars 196 classes)
YOLO cropper: optional · experimental · off by default — full-image is the official prediction
Project: TwinCar · Group 2 · Brainster DS Academy 2026