mlx-swift-AnyCalib
Copyright 2026 Hiroaki Yamane

This product includes software developed as a port of AnyCalib.

--------------------------------------------------------------------------------
AnyCalib
  https://github.com/javrtg/AnyCalib
  Copyright (c) Javier Tirado-Garín and Javier Civera, I3A, University of Zaragoza
  Licensed under the Apache License, Version 2.0.

  "AnyCalib: On-Manifold Learning for Model-Agnostic Single-View Camera
  Calibration", ICCV 2025.

  The model architecture, camera models, on-manifold optimization, and the
  pretrained weights redistributed by this project all originate there. The
  Swift/MLX sources in Sources/MLXAnyCalib are a translation of that PyTorch
  implementation.

--------------------------------------------------------------------------------
DINOv2
  https://github.com/facebookresearch/dinov2
  Copyright (c) Meta Platforms, Inc. and affiliates.
  Licensed under the Apache License, Version 2.0.

  The backbone (Sources/MLXAnyCalib/DINOv2.swift) is a translation of DINOv2's
  vision transformer, vendored into AnyCalib and fine-tuned by its authors.

--------------------------------------------------------------------------------
Modifications made by this project
  * Translated from PyTorch to Swift/MLX.
  * Checkpoint weights are converted from PyTorch .pt to safetensors, with
    convolution tensors transposed from NCHW to NHWC for MLX's native layout.
    No weight values are altered beyond an optional float32 -> float16 cast.
  * Deviations from the reference implementation's numerics are enumerated in
    the "Deliberate deviations" section of README.md.