# Motorcycle Brand Classification App

This app compares 3 image classification approaches on motorcycle images:

- Fine-tuned ViT model [`durovali/vit-motorcycle`](https://huggingface.co/durovali/vit-motorcycle)
- Zero-shot CLIP (`openai/clip-vit-large-patch14`)
- OpenAI vision model (GPT-4.1 Vision)

## Dataset Description

- Custom dataset created manually from web images (Google Images, Unsplash, Pexels)
- Number of classes: `6` (bmw, honda, kawasaki, suzuki, triumph, yamaha)
- Total images: `55` (~9 images per class)
- Split: 80% train (44 images) / 20% validation (11 images)
- Image formats: mixed JPG and PNG
- Dataset loaded using HuggingFace `imagefolder` builder

## Preprocessing Steps

- Images loaded via HuggingFace `imagefolder` dataset builder
- Converted to RGB format
- **Training transforms:** RandomResizedCrop (224px), RandomHorizontalFlip, ToTensor, Normalize
- **Validation transforms:** Resize (224px), CenterCrop (224px), ToTensor, Normalize
- Normalized with ImageNet mean and std values from ViT processor

## Model and Evaluation

- Base model: `google/vit-base-patch16-224-in21k` (Vision Transformer)
- Fine-tuned with transfer learning on custom motorcycle dataset
- Training: 10 epochs, learning rate 2e-5, batch size 16, fp16 on GPU (Tesla T4)
- Hugging Face model: [https://huggingface.co/durovali/vit-motorcycle](https://huggingface.co/durovali/vit-motorcycle)

## Training Performance

| Training Loss | Epoch | Validation Loss | Accuracy |
|---:|---:|---:|---:|
| 1.8418 | 1 | 1.7780 | 0.1818 |
| 1.7629 | 2 | 1.7664 | 0.4545 |
| 1.7311 | 3 | 1.7577 | 0.3636 |
| 1.7177 | 5 | 1.7402 | 0.4545 |
| 1.6922 | 10 | 1.7145 | 0.4545 |

## Example Image Results

The table below reports the true class and Top-3 predictions for ViT and CLIP.

| Image | True Class | ViT Top-3 (score) | CLIP Top-3 (score) | OpenAI (label, confidence) |
|---|---|---|---|---|
| `bmw.jpg` | `bmw` | `honda` (0.2178)<br>`kawasaki` (0.1764)<br>`suzuki` (0.1591) | `bmw` (0.9804)<br>`yamaha` (0.0165)<br>`triumph` (0.0019) | `bmw` (0.95) |
| `honda.jpg` | `honda` | `honda` (0.2035)<br>`kawasaki` (0.1863)<br>`yamaha` (0.1599) | `honda` (0.4927)<br>`yamaha` (0.4869)<br>`suzuki` (0.0100) | `honda` (0.95) |
| `kawasaki.jpg` | `kawasaki` | `kawasaki` (0.2186)<br>`honda` (0.2039)<br>`bmw` (0.1713) | `yamaha` (0.7077)<br>`kawasaki` (0.1124)<br>`bmw` (0.0653) | `kawasaki` (0.95) |
| `triumph.jpg` | `triumph` | `honda` (0.2249)<br>`bmw` (0.1748)<br>`kawasaki` (0.1738) | `triumph` (0.9904)<br>`bmw` (0.0071)<br>`yamaha` (0.0017) | `triumph` (0.98) |
| `yamaha.jpg` | `yamaha` | `honda` (0.1947)<br>`yamaha` (0.1914)<br>`kawasaki` (0.1635) | `yamaha` (0.9057)<br>`bmw` (0.0354)<br>`honda` (0.0213) | `yamaha` (0.90) |

## Comparison Results

| Model | Type | Correct Predictions | Accuracy | Notes |
|---|---|---|---|---|
| ViT (fine-tuned) | Closed-set | 1/5 (kawasaki) | 20% | Too few training images (~9 per class) |
| CLIP | Zero-shot (Open-Source) | 4/5 | 80% | Very strong without any training! |
| OpenAI GPT-4.1 | Zero-shot (Closed-Source) | 5/5 | 100% | Perfect – recognizes logos and design details |

**Key finding:** The fine-tuned ViT model underperforms due to the very small dataset size (~ 55 images total). CLIP and OpenAI perform much better as zero-shot models because they were pre-trained on massive datasets. With more training data (~ 500+ images per class), the fine-tuned model would likely outperform zero-shot approaches.
## Links

- Model: [https://huggingface.co/durovali/vit-motorcycle](https://huggingface.co/durovali/vit-motorcycle)
- App: [https://huggingface.co/spaces/durovali/motorcycle-classification](https://huggingface.co/spaces/durovali/motorcycle-classification)