docs/models.mdx

---
title: "Matching Algorithms"
description: "Choose the right algorithm for your image matching needs"
---

# Matching Algorithms

Image Matching WebUI supports 20+ state-of-the-art algorithms through the [vismatch](https://github.com/gmberton/vismatch) library. Each algorithm is designed for specific use cases and image types.

<Tip>
<strong>Recommendation:</strong> Start with <code>superpoint-lightglue</code> for general-purpose matching. It's fast, accurate, and works well on most image types.
</Tip>


## Algorithm Categories

### 🎯 Sparse Matching (Recommended)

Best for: Feature matching, object detection, real-time applications

| Algorithm | Speed | Accuracy | Best For |
|-----------|-------|----------|----------|
| **superpoint-lightglue** | Fast | High | General purpose |
| **sift-lightglue** | Medium | Very High | High-quality features |
| **disk-lightglue** | Fast | High | Efficient matching |
| **omniglue** | Medium | High | Cross-domain images |
| **aliked** | Very Fast | Medium | Real-time applications |

### 🌊 Dense Matching

Best for: Depth estimation, 3D reconstruction, photogrammetry

| Algorithm | Speed | Accuracy | Memory |
|-----------|-------|----------|--------|
| **roma** | Slow | Very High | High |
| **tiny-roma** | Medium | High | Medium |
| **loftr** | Medium | Very High | High |
| **xoftr** | Slow | Very High | Very High |
| **duster** | Fast | Medium | Low |

### ⚡ Semi-Dense

Best for: Balance between speed and accuracy

| Algorithm | Speed | Accuracy | Use Case |
|-----------|-------|----------|----------|
| **aspanformer** | Slow | Very High | Large baselines |
| **matchformer** | Medium | High | General matching |
| **eloftr** | Fast | High | Real-time matching |


## Algorithm Selection Guide

<CardGroup cols={2}>
  <Card title="Speed Priority" icon="zap">
    **aliked** or **disk-lightglue** for fastest results
  </Card>
  <Card title="Accuracy Priority" icon="check-square">
    **roma** or **superpoint-lightglue** for best quality
  </Card>
  <Card title="Low Memory" icon="hard-drive">
    **duster** or **aliked** for limited GPU memory
  </Card>
  <Card title="General Purpose" icon="star">
    **superpoint-lightglue** works well on most tasks
  </Card>
</CardGroup>


## Popular Algorithms

### SuperPoint + LightGlue

**Best for**: General-purpose image matching

Select `superpoint-lightglue` from the matcher dropdown in the web interface.

**Advantages**:
- Fast and accurate
- Works with various image types
- Robust to scale and rotation changes

### LoFTR

**Best for**: Wide-baseline matching, texture-poor images

Select `loftr` from the matcher dropdown in the web interface.

**Advantages**:
- Excellent for large viewpoint changes
- Works well with low-texture images
- Dense matching for comprehensive coverage

### ROMA

**Best for**: High-precision geometric tasks

Select `roma` from the matcher dropdown in the web interface.

**Advantages**:
- Highest accuracy
- Good for 3D reconstruction
- Robust to extreme viewpoint changes


## Performance Tips

### Speed Optimization

1. **Choose faster algorithms**: `aliked`, `disk-lightglue`
2. **Reduce image resolution**: Set lower resize values
3. **Limit keypoints**: Decrease `max_keypoints`
4. **Use CPU for testing**: Fallback to CPU when GPU unavailable

### Accuracy Optimization

1. **Use state-of-the-art**: `roma`, `superpoint-lightglue`
2. **Enable RANSAC**: Set high confidence levels
3. **Higher thresholds**: Adjust `setting_threshold`
4. **Preprocessing**: Grayscale conversion often improves results

### Memory Management

1. **Use semi-dense algorithms**: `loftr`, `aspanformer`
2. **Reduce batch size**: Process images sequentially
3. **Lower resolution**: Resize images before matching


## Algorithm Comparison

<AccordionGroup defaultOpenAll>
  <Accordion title="Sparse vs Dense vs Semi-Dense">
    <div class="space-y-4">
      <div>
        <div class="font-semibold">Sparse</div>
        <div class="">Fast, memory-efficient, good for feature-based tasks</div>
      </div>
      <div>
        <div class="font-semibold">Dense</div>
        <div class="">Comprehensive matching, high accuracy, memory-intensive</div>
      </div>
      <div>
        <div class="font-semibold">Semi-Dense</div>
        <div class="">Balanced speed and accuracy, good for most applications</div>
      </div>
    </div>
  </Accordion>
  <Accordion title="When to Use Each Category">
    <div class="space-y-2">
      <div><strong>Sparse:</strong> Feature matching, object detection, real-time</div>
      <div><strong>Dense:</strong> 3D reconstruction, depth estimation, photogrammetry</div>
      <div><strong>Semi-Dense:</strong> General matching, medical imaging, document processing</div>
    </div>
  </Accordion>
</AccordionGroup>


## Getting Available Algorithms

List all available algorithms via Python:

```python
from imcui.ui import get_available_model_names
print(get_available_model_names())
```

<Note>
<strong>New algorithms:</strong> All matching algorithms are maintained in the <a href="https://github.com/gmberton/vismatch">vismatch</a> repository. New algorithms become automatically available when added to the collection.
</Note>