---
title: "Configuration"
description: "Customize Image Matching WebUI behavior with YAML configuration"
---

# Configuration

Image Matching WebUI uses YAML configuration files to customize behavior and default parameters.


## Configuration File Location

Configuration files are loaded in priority order (first found):

1. **Custom path**: `-c /path/to/config.yaml`
2. **Working directory**: `./app.yaml`
3. **Config subdirectory**: `./config/app.yaml`
4. **Package default**: `imcui/config/app.yaml`


## Essential Settings

### Device Selection

Device is automatically detected at runtime based on your hardware:

- **CUDA**: Auto-selected if CUDA GPU is available
- **CPU**: Default fallback when CUDA is unavailable
- **MPS**: Auto-selected on macOS with Apple Silicon

<AccordionGroup>
  <Accordion title="Force CPU Mode">
    Set environment variable before running:
    ```bash
    CUDA_VISIBLE_DEVICES="" imcui
    ```
  </Accordion>
  <Accordion title="Select Specific GPU">
    Use environment variable:
    ```bash
    export CUDA_VISIBLE_DEVICES=0
    imcui
    ```
  </Accordion>
</AccordionGroup>

### Default Matcher

There is no `default_matcher` configuration. Matchers are loaded dynamically from the [vismatch](https://github.com/gmberton/vismatch) package and selected through the UI dropdown during matching.

Available matchers: See <a href="/docs/models">algorithm reference</a>


## Matching Parameters

### Thresholds

```yaml
defaults:
  setting_threshold: 0.1      # Detection sensitivity
  keypoint_threshold: 0.05    # Keypoint detection
  match_threshold: 0.2        # Match confidence
  max_keypoints: 2000         # Maximum feature points
```

### RANSAC Filtering

```yaml
defaults:
  enable_ransac: true
  ransac_method: CV2_USAC_MAGSAC
  ransac_reproj_threshold: 8.0
  ransac_confidence: 0.9999
  ransac_max_iter: 10000
  ransac_num_samples: 4
```

Supported RANSAC methods:
- `RANSAC` - Basic random sampling
- `CV2_USAC_MAGSAC` - Magnitude-consistent RANSAC (recommended)
- `LMEDS` - Least Median of Squares

### Geometry Estimation

```yaml
defaults:
  setting_geometry: Homography    # Options: Homography, Fundamental, Essential
```

<AccordionGroup>
  <Accordion title="Essential Matrix Settings">
    Note: Essential matrix parameters are configured through the UI when selecting Essential matrix geometry type.
    Requires camera intrinsics (focal length and principal point).
  </Accordion>
</AccordionGroup>


## Example Configurations

### Minimal Setup

```yaml
defaults:
  setting_threshold: 0.1
  max_keypoints: 2000
  enable_ransac: true
```

### Production Configuration

```yaml
defaults:
  setting_threshold: 0.1
  keypoint_threshold: 0.05
  match_threshold: 0.2
  max_keypoints: 5000
  enable_ransac: true
  ransac_method: CV2_USAC_MAGSAC
  ransac_reproj_threshold: 8.0
  ransac_confidence: 0.9999
  ransac_max_iter: 10000
  setting_geometry: Homography
```

### Development Configuration

```yaml
defaults:
  setting_threshold: 0.2
  max_keypoints: 500
  match_threshold: 0.3
  enable_ransac: false
```

<Note>
<strong>Developer note:</strong> Set <code>CUDA_VISIBLE_DEVICES=""</code> before running to force CPU mode.
</Note>


## Advanced Configuration

### Dataset Path

```yaml
example_data_root: /path/to/custom/datasets
```

<Note>
<strong>Tip:</strong> Dataset paths can also be set via environment variable (`IMCUI_DATA_DIR`) or CLI flag (`-d`).
</Note>


## Configuration Profiles

Create multiple configurations for different use cases:

```bash
# Fast matching for previewing
imcui -c config_fast.yaml

# High-quality matching for production
imcui -c config_high_quality.yaml

# CPU-only matching for testing
CUDA_VISIBLE_DEVICES="" imcui -c config_cpu.yaml -d /test/data
```


## Using Matchers

Matchers are selected through the web interface dropdown. The available matchers are automatically loaded from the vismatch package:

- **Sparse matchers**: SuperPoint + LightGlue, ORB, SIFT, etc.
- **Dense matchers**: LoFTR, RoMa, etc.

<Note>
<strong>Need help selecting a matcher?</strong> Check our <a href="/docs/models">algorithm guide</a> for performance characteristics and recommendations.
</Note>