Datasets:
Update README.md
Browse files
README.md
CHANGED
|
@@ -29,7 +29,6 @@ We categorize the evaluation into **four main task families** to support diverse
|
|
| 29 |
- 📝 **Image Captioning (IC)**: Requires concise, accurate descriptions for both **Global** scenes (whole-image summary) and **Regional** details (directional sectors)。
|
| 30 |
- 🔍 **Single-Image Evaluation (SIE)**: A specialized protocol to test deep understanding of ultra-high-resolution images (4K to $3 \times 10^8$ pixels), probing multi-scale perception and reasoning on a per-image basis.
|
| 31 |
|
| 32 |
-
|
| 33 |
|
| 34 |
|
| 35 |
## 🔖Evaluation Results
|
|
@@ -47,11 +46,25 @@ Closed-source models dominate the leaderboard, yet they still struggle with comp
|
|
| 47 |
|
| 48 |
We further analyze the correlation between perception and reasoning capabilities using Open-Ended VQA evaluation to avoid random guessing.
|
| 49 |
|
| 50 |
-
|
| 51 |
|
| 52 |
|
| 53 |
### 📏 3. Impact of Resolution (Key Insight)
|
| 54 |
|
| 55 |
Does higher resolution support lead to better performance? Our Single-Image Evaluation reveals a critical robustness issue.
|
| 56 |
|
| 57 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
- 📝 **Image Captioning (IC)**: Requires concise, accurate descriptions for both **Global** scenes (whole-image summary) and **Regional** details (directional sectors)。
|
| 30 |
- 🔍 **Single-Image Evaluation (SIE)**: A specialized protocol to test deep understanding of ultra-high-resolution images (4K to $3 \times 10^8$ pixels), probing multi-scale perception and reasoning on a per-image basis.
|
| 31 |
|
|
|
|
| 32 |
|
| 33 |
|
| 34 |
## 🔖Evaluation Results
|
|
|
|
| 46 |
|
| 47 |
We further analyze the correlation between perception and reasoning capabilities using Open-Ended VQA evaluation to avoid random guessing.
|
| 48 |
|
|
|
|
| 49 |
|
| 50 |
|
| 51 |
### 📏 3. Impact of Resolution (Key Insight)
|
| 52 |
|
| 53 |
Does higher resolution support lead to better performance? Our Single-Image Evaluation reveals a critical robustness issue.
|
| 54 |
|
| 55 |
+
|
| 56 |
+
<table>
|
| 57 |
+
<tr>
|
| 58 |
+
<td align="center">
|
| 59 |
+
<img src="assets/image_201.png" width="600">
|
| 60 |
+
</td>
|
| 61 |
+
<td align="center">
|
| 62 |
+
<img src="assets/image_204.png" width="400">
|
| 63 |
+
</td>
|
| 64 |
+
<td align="center">
|
| 65 |
+
<img src="assets/image_205.png" width="400">
|
| 66 |
+
</td>
|
| 67 |
+
</tr>
|
| 68 |
+
</table>
|
| 69 |
+
|
| 70 |
+
|