DINOv2-base Fine-tuned on FGVC-Aircraft (100 Variants)

This model classifies aircraft photos into 100 fine-grained variant classes (e.g. 737-800, A320, Cessna 172). It is part of the Aviation Intelligence System — a ZHAW semester project.

Model Details

  • Backbone: facebook/dinov2-base (86M params, self-supervised pre-training)
  • Task: 100-class fine-grained image classification
  • Training data: FGVC-Aircraft (6,700 images) + Wikimedia Commons extras (2,001 images) = 8,701 total
  • Training: 20 epochs, AdamW (lr=5e-5), cosine schedule, label smoothing (0.1), RandAugment + RandomErasing
  • Hardware: Google Colab T4 GPU, ~60 min

Performance

Metric Value
Top-1 accuracy 84.5%
Top-5 accuracy 97.0%
Macro-average F1 0.84

Evaluated on the FGVC-Aircraft test split (3,333 images).

Comparison

Model Top-1 Top-5
This model (DINOv2-base) 84.5% 97.0%
CLIP zero-shot (no training) 32.8% 77.4%

Usage

from transformers import pipeline

classifier = pipeline("image-classification", model="dubattim/aviation-intelligence-vit-fgvc")
results = classifier("photo.jpg", top_k=5)
print(results)
# [{'label': 'A380', 'score': 0.78}, {'label': 'A340-600', 'score': 0.05}, ...]

Classes (100)

The model recognizes the 100 FGVC-Aircraft variants: 707-320, 727-200, 737-200, 737-300, 737-400, 737-500, 737-600, 737-700, 737-800, 737-900, 747-100, 747-200, 747-300, 747-400, 757-200, 757-300, 767-200, 767-300, 767-400, 777-200, 777-300, A300B4, A310, A318, A319, A320, A321, A330-200, A330-300, A340-200, A340-300, A340-500, A340-600, A380, ATR-42, ATR-72, An-12, BAE 146-200, BAE 146-300, BAE-125, Beechcraft 1900, Boeing 717, C-130, C-47, CRJ-200, CRJ-700, CRJ-900, Cessna 172, Cessna 208, Cessna 525, Cessna 560, Challenger 600, DC-10, DC-3, DC-6, DC-8, DC-9-30, DH-82, DHC-1, DHC-6, DHC-8-100, DHC-8-300, DR-400, Dornier 328, E-170, E-190, E-195, EMB-120, ERJ 135, ERJ 145, Embraer Legacy 600, Eurofighter Typhoon, F-16A/B, F/A-18, Falcon 2000, Falcon 900, Fokker 100, Fokker 50, Fokker 70, Global Express, Gulfstream IV, Gulfstream V, Hawk T1, Il-76, L-1011, MD-11, MD-80, MD-87, MD-90, Metroliner, Model B200, PA-28, SR-20, Saab 2000, Saab 340, Spitfire, Tornado, Tu-134, Tu-154, Yak-42.

Links

License

MIT

Downloads last month
11
Safetensors
Model size
86.7M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Dataset used to train dubattim/aviation-intelligence-vit-fgvc

Space using dubattim/aviation-intelligence-vit-fgvc 1

Evaluation results

  • Top-1 Accuracy on FGVC-Aircraft (100 variants)
    self-reported
    0.845
  • Top-5 Accuracy on FGVC-Aircraft (100 variants)
    self-reported
    0.970