OpenCV

What's new in this version:

Core module:
- Modified Input/OutputArray methods to handle 'std::vector' or 'std::vector<std::vector>' in more accurate way
- Made cuda::GpuMatND compatible with InputArray/OutputArray
- Forced output type for empty matrices where it's defined in API
- Added std::vector length check Input/OutputArray
- Added 16-bit LUT and corresponding HAL entrypoint
- Add cv::Mat::copyAt for for ROI operation
- Extended JSON support in cv::FileStorage: null parsing
- Added support parsing null in JSON parser in cv::FileStorage
- Fixed cv::solveCubic numerical instability via coefficient normalization
- Fixed tempfile race condition on Windows
- Restore parallel framework name on failure attempt
- Dropped OPENCV_FOR_OPENMP_DYNAMIC_DISABLE environment variable in favor of standard OMP_DYNAMIC
- Enabled fp16 conversions, but disabled NEON FP16 arithmetics on Windows for ARM
- Fixed dot product accumulation causing NORM test failures on Windows ARM64
- Added several performance optimizations for Windows on ARM:
- Fixed missing vec_cvfo on IBM POWER9 due to unavailable VSX float64 conversion
- Fix in-place flip in HAL for RISC-V RVV
- Added inRange HAL entry point
- Disabled IPP with AVX512 in cv::compare because of performance regression

Imgproc module:
- Added Iterative Phase Correlation
- Added cv::minEnclosingConvexPolygon
- Improved near-zero convexity handling in cv::convexHull
- Kept cv::convexHull output indices monotone if possible
- Fixed overflow in connected components
- Corrected cv::minAreaRect to follow documentation (force angle to range [-90, 0))
- Increased minAreaRect accuracy with doubles
- Added cv::CLAHE::BitShift option
- Fixed cv::minEnclosingCircle
- Fixed standard cv::HoughLines output shift for rho.
- Fixed LINE_4/LINE_8 swap in cv::drawContours
- Fixed bug in approxPolyDP: calculate distance to a segment, not to a line
- Various performance optimizations for Gaussian Blur, including AVX512
- Added CV_SIMD_SCALABLE branch to pointSetBoundingRect
- Added Canny, Scharr and Sobel for riscv-rvv hal
- Optimized ThickLine() for distant points
- Fixed heap-buffer-overflow in stackBlur
- Fixed the out-of-bounds read in cv::bilateralFilter for 32f images
- Fixed memory leak in ipp_warp function
- Set limitation to IPP Bilateral Filter tiles number to avoid too small tiles
- Fixed incorrect implementation of multi-threaded call for warp perspective in IPP HAL
- Added stateless HAL for filters and morphology

Imgcodecs module:
- Extended metadata support in Jpeg (ICCP and XMP)
- Added OpenEXR multispectral read/write support
- Added strict validation for encoding parameters
- Relaxed decoding size limit to over 1 GiB for several formats
- Added IMWRITE_PNG_ZLIBBUFFER_SIZE option support for PNG
- Fixed memory leak in WebP backend
- Added workaround for image flipping bug in older GDAL FITS drivers
- Added support for 32bpp BMP with BI_BITFIELDS
- Fixed several overflow and out-of-bound access issues in BMP
- Fixed potential crashes in PNG found by fuzzer. Namely 429429085, 429645379, 430091585
- Fixed several potential overflows and error handling mistakes in PNG
- Fixed alpha handling for blending in PNG format
- Set matrixCoefficients to UNSPECIFIED for monochrome images in AVIF
- Added safety checks for AVIF decoder and encoder

VideoIO module:
- Added option to open camera device by index through FFmpeg
- Extended Orbbec SDK support: added timestamps
- Extended Aravis SDK support: allowed system-wide installation of SDK
- Added macOS support for Orbbec Gemini330 camera
- Fixed frame seeking with negative DTS values in FFMPEG backend
- Fixed CvCapture_FFMPEG::setProperty(CAP_PROP_POS_*) followed by getProperty
- Added Raspberry Pi 4 and 5 V4L2 Stateless HEVC Hardware Acceleration with FFmpeg
- Optimized FFmpeg VideoCapture with swscale threads option
- Fix return code from CvVideoWriter_FFMPEG::writeFrame() when encapsulating
- Added OPENCV_FFMPEG_SKIP_LOG_CALLBACK to preserve custom FFmpeg logging
- Get code to compile without FFMPEG's libavdevice
- catch _com_error exceptions to suppress debugger flooding
- Fixed picture_sw object leak in ffmpeg backend with hardware codecs.
- videoio(dshow): Fix incorrect mapping for White Balance Temperature (CAP_PROP_WB_TEMPERATURE) property
- Fixes a crash on iOS when calling VideoWriter::release() in OpenCV 4.12.0

Features2D module:
- Fixed ORB inconsistency for masks with values differs from 255
- Fixed pointer arithmetic in BRISK corner traversal
- Added heuristic to allocate buffer for FAST features depending on input image resolution
- Optimized AGAST for Windows on ARM
- Added FAST to RISC-V RVV HAL

Objdetect module:
- Added pixel-based confidence in ArUco marker detection
- Improved multiple QR codes detection with QRCodeDetector::detectAndDecodeMulti
- Fixed to prevent QR code decoding from throwing on degenerate source points
- Fix QRCodeDetector::detectAndDecode crash in case of inconsistent detection
- Remove floating point arithmetic from angle computation in QR codes
- Speedup ChArUco by avoiding temporary copies
- Fixed stack-use-after-scope errors in charuco detector
- Used fast-path with INTER_LINEAR resize when ALGO_HINT_APPROX is used in HOG 27706

DNN module:
- Added more layers support: ONNX randomNormalLike
- Allowed protobuffer message to be compiled with LITE_RUNTIME
- Fixed ONNX Convolution node parsing when kernel size is implicit
- Fixed Windograd flag handling in convolution layer
- Added NEON intrinsics implementation of fastGEMM1T function
- Add SVE optimized fastGEMM1T function and SVE dispatch
- Fixed heap-buffer-overflow in DNN NaryEltwiseLayer
- Improved performance of softmax_3d with loop unrolling

Calib3d module:
- Added estimateTranslation2D()
- Replaced Gao P3P with Ding P3P algorithm
- Added QR decomposition option to stereoCalibrate
- Improved fisheye::undistortPoints convergence
- Fixed invalid memory access in USAC
- Fixed out of bounds in calibrateAndNormalizePointsPnP

Stitching module:
- Fixed multiband blender memory leak 27333
- Pass warp parameters by value to avoid CUDA constant races
- Optimized fast.cpp on Windows ARM with loop unrolling

HighGUI module:
- Fixed linking for HighGUI against Qt 6.9 and newer
- Fixed Qt HighGUI lifecycle issue when external QApplication exists
- Fixed memory leak in CvWindow Qt back-end

G-API module:
- Added G-API custom stream sources in Python
- Implemented cfgEnsureNamedTensors option to OpenVINO Params
- Implemented cfgClampOutputs option to OpenVINO Params
- Renamed WorkloadType::notify() -> WorkloadType::set()
- Added support to set workload type dynamically in both OpenVINO and ONNX OVEP
- Fixed null pointer dereference in G-API stateful kernels
- Updated deprecated ov::element::undefined

Video module:
- Added multichannel support for findTransformECC and computeECC
- Add optional template mask for findTransformECC
- Added optional Known Foreground Mask for Background Subtractors
- Add setCoarsestScale to DISOpticalFlow
- Fixed potential pointer overflow in Farnback optical flow

ML module:
- Fixed ml::KDTree::findNearest

Photo module:
- Added 16U and 32F support in merge functions in photo module

JavaScript bindings:
- Added option to wrap opencv_contrib into JS too
- Fixed string property bindings in JS generator
- Update helpers.js to include .delete
- Added deep copy behavior for Mat.clone()
- Removed deprecated DEMANGLE_SUPPORT option
- Tuned CONV_NR_FP32 size for WASM
- Added large scalar kernel for fastconv
- Fixed generation when returned class and function are in a namespace
- Fixed JS bindings for namespaced Ptr factory return types

Python bindings:
- Tuned Python bindings for logging
- Marked Feature2D.detectAndCompute mask as optional in Python type stubs
- Marked distCoeffs/R/D as optional in calib3d functions
- Added DLPACK support
- Fixed memory leaks in pybindings
- Add PathLike type hints for args annotated with CV_WRAP_FILE_PATH
- Fixed memory leak in pyopencv_to for path-like objects
- Tunned Python bindings: use export_name as resolved typename for AliasTypeNode
- python: fix pytest-cov false warnings by using compile() with full file path in exec_file_wrapper
- Fixed python scalar typing issues
- Fixed bug in ifdef state machine in header parser for bindings
- Wrap GridType field in cv::findCirclesGrid parameters structure

Java bindings:
- Introduced option to generate Java code with new Cleaners interface or old finalize()
- Enabled Java wrapper generation for Vec4i
- Added Java wrapper support for List<List>

    Applications and Tools:
- Improved interactive calibration tool: added option to select camera backend
- Aligned calibration tools and samples output fields to make them reusable in further steps
- Used calibration board imperfectness correction only for the cases where it's applicable
- Added option to define Aruco index offset in pattern generator
- Moved pattern generator to apps and rewrote tutorial
- Fix charuco_board_pattern in generate_pattern.py
- Updated Tutorial PSNR for identical images

Build:
- Integrated KleidiCV version 0.7 with Mac and Linux support (enabled by default)
- Integrated IPP ICV with AWX512 support.
- Fixed issues identified by PVS Studio
- Added support MS Visual Studio 2026 support
- Added CUDA 13.0 support
- Fixed OpenBLAS detection
- Added reproducible builds support by making host system version optional
- Fixed build issue with some old GCC versions
- Fixed unicode tracing symbols with QT
- Fixed eigen diagnostics for versions after v3.4.0
- Added CMake define option to enforce IPP calls in IPP HAL