[ViralBShah]

Moves OpenCV.jl to OpenCV 5.1 (released 2026-11-06) as version 5.0.0, and wraps ArUco/ChArUco (closes #42) along with many more modules.

Needs OpenCV_jll / OpenCV_CUDA_jll 5.1.0 from JuliaPackaging/Yggdrasil#15577, which builds libopencv_julia from this branch's gen/. CI here will fail until those are registered.

What's wrapped

1,679 Julia functions, up from 602.

• ArUco/ChArUco (OpenCV.aruco): dictionaries, DetectorParameters, ArucoDetector, boards, CharucoDetector, drawing and marker generation.
• Built but not wrapped before: video (optical flow, trackers), stitching, and the rest of objdetect (QR, barcode, FaceDetectorYN, MCC).
• New modules: photo, ml, ximgproc, xphoto, tracking, optflow, and xfeatures2d. AKAZE, KAZE, BRISK and Agast moved to xfeatures2d in 5.x.
• New in 5.1: dnn.Tokenizer (Hugging Face BPE tokenizers), Net KV-cache control, and the XFeat/DISK/ALIKED/LightGlue learned features.

Breaking

• DNN importers: OpenCV 5 removed the TensorFlow, Caffe, Darknet and Torch importers; only ONNX and TFLite remain. logPolar, linearPolar and convertFp16 are gone too.
• Module moves: calib3d is split into calib/geometry/stereo, and features2d is renamed features. Functions keep their names; AKAZE and friends are now OpenCV.xfeatures2d.*.
• Mat type encoding: OpenCV 5 changed it (CV_CN_SHIFT 3 → 5, at most 128 channels). Mats also gain Float16, Bool, UInt32, UInt64 and Int64 element types.

Generator

OpenCV 5 needed generator changes, and several fix bugs that predate 5.x. Details are in gen/PROVENANCE.md:

• Module table: gen_all.py maps each module to the headers it parses, from both opencv and opencv_contrib. regenerate.jl now fetches contrib too.
• Default values: every non-literal default resolves. Unmapped defaults used to become () or Int64(), or drop the keyword method altogether.
• Writable properties: setproperty! now works for CV_PROP_RW fields. It used to cover only the read-only ones, calling setters that don't exist.
• cv::Ptr classes: classes handed out by cv::Ptr factories (Stitcher, the trackers, ml::TrainData) are now held by Ptr, and Algorithm constructors return cv::Ptr. Factory results and constructed objects now work with every method.
• Nested types: nested types and namespaced bases resolve.
• Compiler limit: the C++ module body is split into ~250-line functions. As one function, it made GCC 10 run out of memory.
• MatShape and small Mats: cv::MatShape is now a class, so the four dnn::Net methods that take it are wrapped by hand. 0-D and 1-D Mats convert correctly.

Testing

All of this was done locally on Linux x86_64 with OpenCV_jll/OpenCV_CUDA_jll 5.1.0 built and deployed from that recipe:

• Julia 1.13 (CPU): 802 tests pass, 2 skipped (model-heavy learned features).
• Julia 1.10 (CPU): same result.
• Julia 1.13 with OpenCV_CUDA_jll: same result, with OpenCV bound to the CUDA build.
• Doctests and the docs @repl block pass.
• Regeneration produces identical output across Python hash seeds.

CUDA caveat. OpenCV 5.1's DNN only has the new engine. It runs a graph on the GPU only if every op has a CUDA kernel, and it inserts its own TransformLayout op, which has none. As a result, YuNet and ResNet-50 run entirely on the CPU even with DNN_BACKEND_CUDA; the log says "has no CUDA exec; running the whole graph on CPU". The CUDA build loads, finds CUDA 13 and cuDNN 9.28, and gives identical results, but the ~5× GPU speedup we had with 4.14 is gone. This is an upstream limitation. ENGINE_ORT (ONNX Runtime) could be a later path to GPU inference.

Not covered: macOS, Windows and aarch64 builds; the Yggdrasil CI will be the first to build them. The new modules have smoke tests only.

🤖 Generated with Claude Code