Source, footage & limits

This experiment uses the real SVO Pro odometry implementation through Jianxff's ROS-free WebAssembly port. It does not substitute 2D image motion for a 3D SVO pose.

• Pinned upstream source · Matching upstream source archive · GPLv3 · Artifact hashes and modifications.
• Application built locally; the upstream SVO WebAssembly artifact was reused, not recompiled locally. The appended camera wrapper was replaced; the compiled WASM payload is unchanged.
• Forster, Pizzoli, Scaramuzza. SVO: Fast Semi-Direct Monocular Visual Odometry. ICRA 2014.
• TUM RGB-D benchmark: desk (613 RGB frames), wooden floor (1242 RGB frames). Sturm, Engelhard, Endres, Burgard, Cremers, IROS 2012. CC BY 4.0. RGB rectified using published Freiburg1 calibration, resized to 480×360, saved as JPEG. Depth and ground truth are not provided to the tracker.

Download this application source (GPLv3), including reproducible build, vendoring, sample preparation and tests.

What runs where

SVO reference: CPU WebAssembly worker. WebGPU experiment: GPU grayscale, three-level image pyramid, Shi–Tomasi grid selection and forward/backward pyramidal Lucas–Kanade patch alignment. GPU mode does not yet estimate a 3D pose, map depth, detect a floor plane or anchor an object. Diagnostic patch coordinates are read back for overlay and testing.

The live camera uses approximate intrinsics. No IMU fusion, metric scale, persistent anchors, relocalization guarantee or iPhone performance has been established. A successful desktop dataset replay is not evidence of reliable mobile AR.

Back to the experiment