Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,86 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language: en
|
| 3 |
+
license: apache-2.0
|
| 4 |
+
tags:
|
| 5 |
+
- image-classification
|
| 6 |
+
- computer-vision
|
| 7 |
+
- geolocation
|
| 8 |
+
- geoguessr
|
| 9 |
+
- convnext
|
| 10 |
+
- convnextv2
|
| 11 |
+
datasets:
|
| 12 |
+
- marcelomoreno26/geoguessr
|
| 13 |
+
metrics:
|
| 14 |
+
- accuracy
|
| 15 |
+
- f1
|
| 16 |
+
base_model: facebook/convnextv2-base-22k-224
|
| 17 |
+
model-index:
|
| 18 |
+
- name: convnextv2-geoguessr
|
| 19 |
+
results:
|
| 20 |
+
- task:
|
| 21 |
+
type: image-classification
|
| 22 |
+
name: Image Classification
|
| 23 |
+
dataset:
|
| 24 |
+
type: marcelomoreno26/geoguessr
|
| 25 |
+
name: GeoGuessr
|
| 26 |
+
metrics:
|
| 27 |
+
- type: accuracy
|
| 28 |
+
value: 0.6103
|
| 29 |
+
name: Accuracy
|
| 30 |
+
- type: f1
|
| 31 |
+
value: 0.5177
|
| 32 |
+
name: F1-Macro
|
| 33 |
+
---
|
| 34 |
+
|
| 35 |
+
# ConvNeXt V2 Base Fine-tuned on GeoGuessr
|
| 36 |
+
|
| 37 |
+
## Model Description
|
| 38 |
+
|
| 39 |
+
This model is a **ConvNeXt V2 Base** fine-tuned for **geographic image classification** of Google Street View images. The model can identify the country of origin of an image among **55 different countries**.
|
| 40 |
+
|
| 41 |
+
ConvNeXt V2 is a modern evolution of convolutional networks that incorporates improvements inspired by Vision Transformers, while maintaining the efficiency and spatial inductive bias of traditional CNNs.
|
| 42 |
+
|
| 43 |
+
## Results
|
| 44 |
+
|
| 45 |
+
The model has been trained and evaluated on the GeoGuessr dataset with the following results:
|
| 46 |
+
|
| 47 |
+
| Metric | Value |
|
| 48 |
+
|--------|-------|
|
| 49 |
+
| **Accuracy** | **61.03%** |
|
| 50 |
+
| **F1-Macro** | **51.77%** |
|
| 51 |
+
|
| 52 |
+
### Comparison with Previous Work
|
| 53 |
+
|
| 54 |
+
| Model | Accuracy | F1-Macro |
|
| 55 |
+
|-------|----------|----------|
|
| 56 |
+
| **ConvNeXt V2 (this model)** | **61.03%** | **51.77%** |
|
| 57 |
+
| SigLIP2 (prithivMLmods) | 64.85% | 38.36% |
|
| 58 |
+
| ViT-Base-384 (dataset author) | 38.81% | 14.40% |
|
| 59 |
+
|
| 60 |
+
Although SigLIP2 has slightly higher accuracy, this model achieves a **significantly better F1-Macro** (+13.41%), indicating more balanced performance across all classes.
|
| 61 |
+
|
| 62 |
+
### Performance by Country
|
| 63 |
+
|
| 64 |
+
**Top 10 best performing countries:**
|
| 65 |
+
- 🇯🇵 Japan: 85.98%
|
| 66 |
+
- 🇧🇼 Botswana: 82.61%
|
| 67 |
+
- 🇩🇪 Germany: 80.18%
|
| 68 |
+
- 🇦🇺 Australia: 78.93%
|
| 69 |
+
- 🇿🇦 South Africa: 78.74%
|
| 70 |
+
- 🇨🇦 Canada: 76.26%
|
| 71 |
+
- 🇬🇧 United Kingdom: 74.06%
|
| 72 |
+
- 🇸🇬 Singapore: 72.63%
|
| 73 |
+
- 🇹🇭 Thailand: 70.34%
|
| 74 |
+
- 🇮🇱 Israel: 69.49%
|
| 75 |
+
|
| 76 |
+
**Top 10 most challenging countries:**
|
| 77 |
+
- 🇱🇻 Latvia: 14.29%
|
| 78 |
+
- 🇸🇰 Slovakia: 18.87%
|
| 79 |
+
- 🇵🇹 Portugal: 21.28%
|
| 80 |
+
- 🇱🇹 Lithuania: 21.43%
|
| 81 |
+
- 🇨🇿 Czechia: 22.38%
|
| 82 |
+
- 🇺🇦 Ukraine: 23.33%
|
| 83 |
+
- 🇭🇷 Croatia: 25.00%
|
| 84 |
+
- 🇩🇰 Denmark: 28.57%
|
| 85 |
+
- 🇧🇪 Belgium: 29.47%
|
| 86 |
+
- 🇮🇹 Italy: 32.23%
|