The same SLAM pipeline that runs on the desktop, compiled to WebAssembly. Features, geometry, mapping and pose come from the identical Rust code, single threaded. Frames are pulled straight out of the file rather than off the video element, so a run sees the same frames whatever else the machine is doing: press Run twice and the digest under the stats is the same both times. Nothing is uploaded anywhere!
docs.rs · source on GitHub · crates.io
Start with a sample the tracker handles well, or upload your own:
Samples are clips from the TUM RGB-D benchmark (Sturm et al., IROS 2012), used under CC BY 4.0, with their calibration filled in.
For an upload, the focal length defaults to 0.85 × width.
Top-down view. The estimate is blue; for a sample, ground truth is green and the estimate is Sim(3) aligned onto it. For an upload, keyframes are the black marks. Hover a point to see the frame it came from.
waiting for a video