This project classifies utility pole images by country using both a fine-tuned ViT model and a zero-shot CLIP model.
| Name | URL |
|---|---|
| Huggingface | Huggingface Space |
| Model Page | Huggingface Model Page |
| Code | GitHub Repository |
The different pole classes correspond to countries, including:
['Albania', 'Argentina', 'Australia', 'Austria', 'Bangladesh', 'Belgium', ..., 'Turkey', 'Uganda', 'Ukraine', 'United', 'Uruguay']
(96 total classes)
| Data Source | Description |
|---|---|
| GeoHints Utility Poles Dataset | All utility pole images were fetched from this dataset, which provides annotated images of utility poles collected from various global sources. |
| Split | % of Total | Approximate Size |
|---|---|---|
| Train | 80% | — |
| Validation | 10% | — |
| Test | 10% | — |
| Epoch | Training Loss | Validation Loss | Accuracy | Precision | Recall | F1 Score |
|---|---|---|---|---|---|---|
| 1 | No log | 4.4879 | 2.60% | 0.17% | 2.60% | 0.31% |
| 5 | 3.9392 | 4.2822 | 11.69% | 11.33% | 11.69% | 9.26% |
| 10 | 2.5045 | 4.1706 | 14.29% | 11.68% | 14.29% | 10.98% |
| 15 | 1.6740 | 4.1136 | 16.88% | 12.85% | 16.88% | 12.67% |
| 20 | 1.3136 | 4.1020 | 18.18% | 13.63% | 18.18% | 13.47% |
Full logs and charts available on Hugging Face TensorBoard:
🔗 TensorBoard Link
| Model/Method | Accuracy | Precision | Recall |
|---|---|---|---|
ViT Fine-Tuned (google/vit-base-patch16-224) |
18.18% | 13.63% | 18.18% |
CLIP Zero-Shot (openai/clip-vit-base-patch32) |
16.0% | 20% | 18% |
Upload an image of a utility pole to see predictions from both models.
App: Huggingface Space
