t1an-xyz/plane-classifier

Commercial Plane Family Classifier

Fine-tuned MobileNetV4 (mobilenetv4_conv_medium) classifying 16 commercial aircraft families from planespotting photos and camera crops.

Trained on the nyuuzyou/aircraft-images dataset with heavy field corruptions (blur, crops, noise, contrast, weather shift) to make predictions robust to phone photos and landing approaches.

Performance

• Pristine Test Accuracy: ~93.2%
• Field-Sim Test Accuracy: ~93.1%
• Macro-F1: ~0.904

Supported Aircraft Families (16 Classes)

1. Airbus A220
2. Airbus A310
3. Airbus A320
4. Airbus A330
5. Airbus A340
6. Airbus A350
7. Airbus A380
8. Boeing 737
9. Boeing 747
10. Boeing 757
11. Boeing 767
12. Boeing 777
13. Boeing 787
14. Bombardier CRJ
15. Embraer E-Jet
16. McDonnell Douglas MD-80/90

Field Test

---

Quickstart (Inference with PyTorch & timm)

import torch
import timm
from torchvision import transforms
from PIL import Image
from huggingface_hub import hf_hub_download

REPO_ID = "t1an-xyz/plane-classifier"

# 1. Download checkpoint from Hugging Face
ckpt_path = hf_hub_download(repo_id=REPO_ID, filename="plane_classifier_checkpoint.pt")
ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=False)

classes = ckpt["classes"]
model_name = ckpt["model_name"]
img_size = ckpt.get("img_size", (256, 256))
norm_mean = ckpt.get("norm_mean", (0.485, 0.456, 0.406))
norm_std = ckpt.get("norm_std", (0.229, 0.224, 0.225))

# 2. Reconstruct timm model & load weights
model = timm.create_model(model_name, pretrained=False, num_classes=len(classes))
model.load_state_dict(ckpt["state_dict"])
model.eval()

# 3. Preprocess image
preprocess = transforms.Compose([
    transforms.Resize(img_size),
    transforms.ToTensor(),
    transforms.Normalize(mean=norm_mean, std=norm_std),
])

img = Image.open("plane.jpg").convert("RGB")
input_tensor = preprocess(img).unsqueeze(0)

# 4. Predict
with torch.no_grad():
    logits = model(input_tensor)
    probs = torch.softmax(logits, dim=1)[0]

top3_indices = torch.topk(probs, k=3).indices.tolist()
for idx in top3_indices:
    print(f"{classes[idx]}: {probs[idx]:.2%}")