COLMAP WebGPU

Photos to a sparse COLMAP model for Gaussian splat training. All processing stays on this device.

Checking WebGPU…

Experimental sparse mapper built with COLMAP's VLFeat, PoseLib and a browser-specific reconstruction controller. This is not the full upstream COLMAP application.

Photos Or folder
Folders include images in subfolders. Choose 2–200 images per reconstruction; other files are ignored.
Calibration JSON
Estimate and refine missing calibration Same camera and zoom
Fallback focal / longest side Image maximum side Features / image

Calibration uses fx, fy, cx, cy in original decoded-image pixels; optional k1, k2, p1, p2 for undistortion. Supply an object keyed by original filenames. Supplied calibration stays fixed. Otherwise, EXIF or the fallback focal value initializes automatic focal and radial-distortion refinement. Uncheck “Same camera and zoom” for mixed cameras without identifying EXIF metadata. Principal points stay at image center for automatic calibration. Use varied viewpoints and scene depth; narrow arcs or planar scenes can leave focal length ambiguous even with low reprojection error.

Extract features
Match & verify pairs
Reconstruct cameras
Final optimization
Prepare export

Ready to begin · 5 stages

Stages show processing order, not equal amounts of time.