COLMAP Reconstruction Plugin for LichtFeld Studio

A sparse reconstruction plugin for LichtFeld Studio. It runs COLMAP via pycolmap on an image folder, builds a sparse SfM model, and reports reconstruction metrics directly in the panel.

Features

• Sparse COLMAP Reconstruction
• Incremental and Global Mapping Modes
• Reprojection Error Metrics in UI

Installation

Manual Installation

git clone https://github.com/shadygm/Lichtfeld-COLMAP-Plugin.git ~/.lichtfeld/plugins/Lichtfeld-COLMAP-Plugin

Usage

GUI

1. Open the COLMAP Reconstruction panel in LichtFeld Studio.
2. Select a folder containing overlapping JPG or PNG images.
3. Leave Normal selected for the default workflow, or switch to Low for a lighter setup.
4. Optional: Open Advanced to adjust matcher, reconstruction mode, camera model, and optimization parameters.
5. Click Run Reconstruction.
6. Monitor the stage, progress bar, and live logs while COLMAP runs.
7. Review the output path and reprojection error statistics when reconstruction completes.

Configuration

Presets

• Normal (Default): Incremental reconstruction, exhaustive matching, Downsample=1x (Full Resolution), Max Features=2048, Max Matches=2048, Exhaustive Block Size=15, BA Max Iterations=50.
• Low: Incremental reconstruction, sequential matching, Downsample=2x (Half Resolution), Max Features=1536, Max Matches=1024, Exhaustive Block Size=10, BA Max Iterations=20.

Downsample Multiplier

The plugin applies a downsample multiplier to input images before reconstruction:

Multiplier	Resolution	Use Case
1x	Full resolution	Best quality, slower processing
2x	Half resolution (50%)	Balanced quality and speed
4x	Quarter resolution (25%)	Faster processing, lower quality
8x	Eighth resolution (12.5%)	Fastest processing, draft quality

Images are pre-resized using high-quality Lanczos resampling before feature extraction. The downsampled images are saved as JPEG (quality 95) in the working directory.

Reconstruction Modes

• Incremental: Default and generally the safest option.
• Global (GLOMAP): Available when supported by the installed pycolmap build.

Matching Modes

• Exhaustive: Matches all image pairs. This is the default and is usually best for small to medium image folders.
• Sequential: Matches neighboring filenames for video-like image sequences. Use sequentially ordered names such as image0001.jpg, image0002.jpg, etc.

Reported Metrics

• Mean Reprojection Error
• Median Reprojection Error
• 90th Percentile Reprojection Error

Output

The plugin writes a standard COLMAP-style dataset layout. If you select an existing images/ folder, the dataset root is its parent directory. Otherwise, the plugin creates a sibling dataset folder named after the selected image folder:

<dataset_root>/
├── images/
└── sparse/

After a successful reconstruction, the plugin loads <dataset_root>/ into LichtFeld as the dataset for training.