Rafael Spring has released OfflineTools COLMAP, a free in browser build of COLMAP that reports faster SIFT extraction than the stock native binary.
The build compiles COLMAP to WebAssembly and executes in the page, so no images are uploaded and nothing is installed. It produces camera poses, sparse point clouds, keypoints, and masks, the same outputs a gaussian splatting trainer consumes as initialization.
On a set of 229 images at 5616x3744, OfflineTools reports SIFT GPU extraction finishing in 54 seconds against 1 minute 41 seconds for stock COLMAP, and SIFT CPU extraction in 1 minute 25 seconds against 1 minute 46 seconds. The host was a Ryzen 7 5800X with 32 GB of RAM and an 8 GB RTX GPU. Both comparisons are OfflineTools' own measurements.
COLMAP 4.2.0 shipped its own browser based viewer for sparse binary reconstructions on September 1. That viewer inspects a reconstruction that already exists. The OfflineTools build performs the reconstruction.
Spring previously founded DotProduct, maker of the Dot3D handheld scanning software, and worked at Google before that. OfflineTools is a separate project from DotProduct.
The app page also carries capture guidance, covering three image minimum coverage per feature, locked focus and exposure, sequential matching for ordered video frames, and vocabulary tree matching for large collections. A practitioner on a machine where COLMAP will not build, or without a CUDA workstation, can now produce poses for a splat training run from a browser tab.
New to Gaussian Splatting? Start here
