# Technical Document Extraction: Accuracy Analysis by Image Size Interval
## Chart Type
Bar chart with grouped segments, comparing accuracy metrics across image size intervals.
## Axes
- **X-axis**: "Average Image Size Interval (width x height)"
Categories:
- (0,25] (Total: 19)
- (25,100] (Total: 139)
- (100,225] (Total: 129)
- (225,400] (Total: 51)
- (400,625] (Total: 39)
- (625,900] (Total: 23)
- **Y-axis**: "Accuracy on I_t (%)"
Scale: 0–80% in 10% increments.
## Legend
- **Position**: Top-right corner.
- **Entries**:
1. `GPT-o3-mini RSPC` (Blue)
2. `GPT-o3-mini KAAR` (Light Blue)
3. `QwQ-32B RSPC` (Purple)
4. `QwQ-32B KAAR` (Pink)
## Data Points & Trends
### (0,25] Interval (Total: 19)
- **GPT-o3-mini RSPC**: 73.7% (Blue)
- **QwQ-32B RSPC**: 42.1% (Purple)
- **GPT-o3-mini KAAR**: 15.8% (Light Blue)
- **QwQ-32B KAAR**: 5.3% (Pink)
**Trend**: GPT-o3-mini RSPC dominates; KAAR methods underperform.
### (25,100] Interval (Total: 139)
- **GPT-o3-mini RSPC**: 48.9% (Blue)
- **QwQ-32B RSPC**: 23.7% (Purple)
- **GPT-o3-mini KAAR**: 11.5% (Light Blue)
- **QwQ-32B KAAR**: 5.0% (Pink)
**Trend**: GPT-o3-mini RSPC maintains lead; KAAR accuracy declines.
### (100,225] Interval (Total: 129)
- **GPT-o3-mini RSPC**: 24.8% (Blue)
- **QwQ-32B RSPC**: 8.5% (Purple)
- **GPT-o3-mini KAAR**: 6.2% (Light Blue)
- **QwQ-32B KAAR**: 4.7% (Pink)
**Trend**: Both models show significant drops; KAAR methods remain low.
### (225,400] Interval (Total: 51)
- **GPT-o3-mini RSPC**: 11.8% (Blue)
- **QwQ-32B RSPC**: 9.8% (Purple)
- **GPT-o3-mini KAAR**: 5.9% (Light Blue)
- **QwQ-32B KAAR**: 2.0% (Pink)
**Trend**: GPT-o3-mini RSPC still outperforms; KAAR accuracy near baseline.
### (400,625] Interval (Total: 39)
- **GPT-o3-mini RSPC**: 5.1% (Blue)
- **QwQ-32B RSPC**: 0% (Purple)
- **GPT-o3-mini KAAR**: 0% (Light Blue)
- **QwQ-32B KAAR**: 0% (Pink)
**Trend**: All methods fail; no accuracy recorded.
### (625,900] Interval (Total: 23)
- **GPT-o3-mini RSPC**: 4.3% (Blue)
- **QwQ-32B RSPC**: 0% (Purple)
- **GPT-o3-mini KAAR**: 0% (Light Blue)
- **QwQ-32B KAAR**: 0% (Pink)
**Trend**: GPT-o3-mini RSPC marginally functional; others fail.
## Key Observations
1. **Model Performance**:
- GPT-o3-mini RSPC consistently outperforms QwQ-32B RSPC across all intervals.
- KAAR methods (both models) show negligible accuracy, especially in larger intervals.
2. **Image Size Impact**:
- Accuracy declines sharply as image size increases.
- GPT-o3-mini RSPC retains ~70% accuracy in the smallest interval but drops to 4.3% in the largest.
3. **KAAR Limitations**:
- KAAR methods fail to achieve meaningful accuracy in intervals larger than (25,100].
## Spatial Grounding
- Legend colors match bar segments exactly (e.g., blue = GPT-o3-mini RSPC).
- No textual data outside the chart; all information embedded in bars and legend.
## Language
- All text in English. No non-English content detected.
