mprib/geocalib-fields-onnx

GeoCalib field net, ONNX export

The convolutional half of GeoCalib (Veicht et al., ECCV 2024, ETH CVG) as a single self-contained ONNX file: image in, four dense perspective fields out. Pair it with a small Levenberg-Marquardt solver on those fields to estimate roll, pitch, and gravity direction from a single image with no torch dependency.

Exported for Caliscope's board-free vertical estimation.

What it is

GeoCalib runs in two stages: a network predicts, per pixel, the projected world-up direction (up_field), the latitude of the viewing ray (latitude_field), and a confidence for each; a Levenberg-Marquardt optimizer then fits camera tilt (and optionally focal) to those fields. This export is the network stage only, from the official pinhole weights. The solver stage is ~300 lines of numpy, shipped with the consumer (Caliscope).

Contract

• Input image: (1, 3, H, W) float32 RGB in [0, 1]. H and W are dynamic but must be multiples of 32. GeoCalib's preprocessing resizes the short side to 320 with bilinear+antialias and rounds edges to multiples of 32; match it for best results.
• Outputs: up_field (1, 2, H, W), up_confidence (1, H, W), latitude_field (1, 1, H, W) (radians), latitude_confidence (1, H, W).
• One graph serves all resolutions and orientations (exported with torch.onnx dynamo, dynamic shapes).
• fp32, 118 MB, weights embedded — no external data file.

Differences from upstream

Upstream GeoCalib is stochastic at inference: its Hamburger/NMF decoder draws random bases on every forward, so identical inputs give slightly different fields run to run (~0.005–0.2° gravity wobble downstream). This export freezes one seeded basis (seed 0) as a graph constant, making it deterministic. Field parity against frozen-basis torch is ~1e-6; gravity parity through the same solver is ~1e-5 degrees — four orders of magnitude below upstream's own run-to-run noise.

Provenance

• Upstream weights: geocalib-pinhole v1.0 release, cvg/GeoCalib at commit 747054d.
• Export procedure: wrap the model's backbone, low-level encoder, and perspective decoder in a module returning the four fields; replace each NMF2D _build_bases with a constant seeded basis (torch.Generator().manual_seed(0), normalized); export with torch.onnx.export(dynamo=True), dynamic height/width, external_data=False. Validated against torch on real frames before publishing.
• SHA-256: a724447e1bf7138352e5e70bbe0f011e6fb02909edd3148a5d0f46e8d153ae7b

License and attribution

Apache 2.0, as a derivative of GeoCalib (© ETH Zurich, Computer Vision and Geometry Group). If you use this model, cite the original work:

@inproceedings{veicht2024geocalib,
  author    = {Alexander Veicht and Paul-Edouard Sarlin and Philipp Lindenberger and Marc Pollefeys},
  title     = {{GeoCalib: Single-image Calibration with Geometric Optimization}},
  booktitle = {ECCV},
  year      = {2024}
}