anycalib-onnx

AnyCalib ONNX

Provider-neutral ONNX exports of the official AnyCalib pinhole camera-intrinsics model for the KitsuMate Vision To Pose Blender extension.

The bundle predicts AnyCalib's dense tangent-coordinate field. The extension applies the pinned image preprocessing and a NumPy port of the official pinhole fit and ten-step Gauss-Newton refinement to recover fx, fy, cx, and cy.

Variants

• onnx/fp16: mixed FP16/FP32 weights and operations, with FP32 input/output boundaries, for validated accelerated execution.
• onnx/fp32: full FP32 graph for CPU and macOS fallback.

Both graphs use ONNX opset 20, accept dynamic NCHW RGB images whose spatial edges are divisible by 14, and output an NCHW two-channel tangent field at input resolution. Exact hashes, sizes, preprocessing contract, upstream revisions, and validation results are recorded in each manifest and parity report.

These models estimate camera intrinsics, not camera roll, pitch, translation, zoom changes, or full 6DoF camera motion.

Provenance and license

The upstream AnyCalib code and checkpoint are Apache-2.0 licensed. AnyCalib incorporates a DINOv2 backbone derived from Meta Platforms' Apache-2.0 implementation. See PROVENANCE.json, NOTICE.md, and LICENSE.