File size: 1,637 Bytes
d9d9996 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 | 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.
|