Previous Change Logs:

• OpenCV v2.2-v4.10
• OpenCV v1.0-v2.1

July, 2025

Summer update for OpenCV 4.x

• Core Module:

  • Added user-defined logger callback #27154
  • Added reinterpret() method to cv::Mat #25394
  • Fixed empty ND-array construction #27362
  • Fixed bug with int64 support for FileStorage #26846
  • Fixed cv::meanStdDev overflow for large images #26867
  • Useed HAL for cv::mean function too #26895
  • Vectorized cv::normalize / cv::norm #26885
  • Enabled SIMD_SCALABLE for exp and sqrt #26886
  • vectorized normDiff with universal intrinsics #27042
  • Refactored normDiff in hal_rvv and extended with support of more data types #27115
  • Further vectorized copyTo with mask #27145
  • Fix POWN OpenCL implementation #27274
  • Restored legacy universal intrinsics operators in a separate header #27327
  • Improved solveCubuc accuracy: #27331, #27347
  • Deprecated copyData Parameter in cv::UMat construction from std::vector and always copy data #27408

• Imgproc Module:

  • Optimized and reduced memory consumption in cv::findContours #26690, #26834
  • Fixed integer overflow in cv::HoughLinesProbabilistic with very large input images #26744
  • Fixed several potential overflows in image warping #26765, #26776
  • Improved robustness for ellipse fitting #26773
  • Added cv::THRESH_DRYRUN flag to get adaptive threshold values without input thresholding #26836
  • Fixed cv::getPerspectiveTransform for singular case #26926
  • Optimized bayer2Gray with universal intrinsics #26868
  • Added optional mask to cv::threshold #26842
  • Accelerated cv::remap calculation with multi-threading, when cv::WARP_INVERSE_MAP is used #27108
  • Reworked HSV color conversion tables initialization in OpenCL branch to fix possible data raise #27287
  • Fixed bilateral filter's sigma color and sigma space issue #27309
  • Impoved performance of medianblur #27299
  • Added cv::getClosestEllipsePoints function to get the closest point on an ellipse #26299
  • Corrected IPP implementation for cv::distanceTransform results with single thread #27432
  • Fixed cv::matchTemplate with mask crash #26703
  • Added diamond structuring element to morphology operations #27441, #27492
  • Improved cv::intersectConvexConvex accuracy in corner cases #26974

• Calib3d module:

  • Fixed inliers updating in homography estimation #26742
  • Added cv::solvePnPRansac implementation for the fisheye camera model #26669
  • Fixed different marker ordering with cv::findChessboardCornersSBWithMeta and cv::CALIB_CB_LARGER flag.
  • Fixed logical mistake in FilterSpecklesImpl #26996
  • Optimized undistort points for the fisheye camera model #26988
  • Optimized new camera matrix estimation and some related functions #27001
  • Added warning if projected axes are out of camera frame in cv::drawAxes function #27311
  • Added image dimension check to avoid StereoSGBM non-determinism #27305
  • Fixed bugs in SQPNP orthogonalization and simplified column vectors copying #27437

• DNN Module:

  • Added TFLite StridedSllice (without strides but just Slice) and enable strides for TF importers #27273
  • Added fusion handler of conv+eltwise in case of multi-output node (i.e. Split) #27326
  • Added new ops import in TFLite parser: SUB, SQRT, DIV, NEG, SQUARED_DIFFERENCE, SUM #27307
  • Fxied scalars and 1d tensors support in TFLite parser #27307
  • Reduced mumber of NHWC<->NCHW conversions in TFLite parser #27307
  • Added checks before getting latest pin in Net::Impl::getLatestLayerPin #27353
  • Added OpenVINO NPU support #27363
  • Added more operations to CANN backend #24756

• Objdetect module:

  • Extend ArUcoDetector to run multiple dictionaries in an efficient manner #26934.
  • Fixed rotated aruco marker board generation #26753
  • Fixed incorrect Aruco marker detection near image edge #26968
  • Fixed invalid vector access in QR de/encoder #26532
  • Fix QR code encoder with auto version #27244
  • Aded QRCode ECI encoding support #24426, #27486
  • Made ChAruco board consistency check optional #26824

• Photo Module:

  • Added 16-bit support to cv::fastNlMeansDenoising and updated tests #26831
  • Local decolor pipeline optimization #27041
  • Fixed Triangle Weights used by Debevec algorithm being non zero at extremes #27270
  • Fixed NaNs in HDR Triangle Weights and Tonemapping and Update LDR Ground Truth in tutorial #27396

• VideoIO Module:

  • Fixed invalid custom FFmpeg options handling provided with environment variable #26793
  • Fixed default cv::CAP_PROP_ORIENTATION_ATUO behaviour #26800
  • Improved camera opening performance when with DShow backed, if custom options are provided #26601
  • Added zoom support into Android native camera capture #26837
  • Fixed V4L frame size for non-BGR output #27193
  • Added Orbbec Gemini 330 camera support #27230
  • Fixed cv::VideoWriter failure on writing colorless images #27153
  • Fixed Android setCameraIndex issue #27419

• Imgcodecs:

  • Extended image I/O API with metadata support #27499
  • Added in-memory animation encoding and decoding #27013
  • Added hidden frame support for Animated PNG #27127
  • Added GIF decode and encode for imgcodecs #25691, #26971, #26859, #26738, #26930, #27081, #27138, #27296
  • Added animated WebP Support #25608, #27457
  • Added GDAL multi-channel support #27458
  • Fixed AVIF linkage on Windows #26762
  • Significantly improved PNG and Animated PNG files handing #26849, #26739, #26748, #26782, #26832, #26835, #26848, #26851, #26854, #26915, #27040, #27087, #27297
  • Added several fixes for libspng integration #27314
  • Fixed memory leak in JpegXL decoder #26787
  • Added support for cv::IMREAD_UNCHANGED and other ImreadFlags into JpegXL implementation #26788
  • Implemented imdecode() for JpegXL to directly read from memory #26844
  • Added lossless compression support for JpegXL #27384
  • Fixed cv::imread() implementation for 16-bit png with libspng back-end #27113

• Highgui Module:

  • displayOverlay doesn't disappear after timeout #27082
  • Fix closing of windows when using the Qt backend #27170
  • Fixed QT window resize regression #27282

• G-API module:

  • Workaround for successfull append of OpenVINO Execution Provider: Moved creation of 'Ort::Env' before it #26983
  • Adding AddRgbFeature(), and improving robustness in ComputeRgbDistance #26682
  • Added MS Media Foundation availability check in G-API too #27355

• Video Module:

  • Added trackers factory with pre-loaded dnn models #26875

• HAL:

  • Extracted HAL implementations as dedicated folder #27252.
  • Implemented new HAL for RISC-V RVV 1.0 platforms. See details in blogpost. See all changes by query
  • Refactored OpenVX implementation as HAL #26903, #26911, #26907.
  • Partially refactored Intel IPP integration as HAL #26880, #27125, #27128, #27194, #27213, #27346. Further IPP conversion will be continued in the next releases
  • Fixed HALs dependency from thre cv namespace #26878
  • Updated KleidiCV HAL for ARM to version 0.5
  • Extended FastCV HAL for Qualcomm platforms #27217, #27216, #27184, #27182, #27033
  • Added new HAL entries for cv::sum #27346, copyTo with mask #27162, cv::DFT and cv::DCT variations #26865, convert with scale #26804, cv::norm variations #26804, minMaxIdx with step support #26789, cv::calcHist #27332, cv::pyrUp #26958, more variations for cv::remap #27119

• Python, Java and JavaScript bindings:

  • Added conditional inclusion support to header parser #27325, #27393, #27478
  • Updated predefined_types.py #26826
  • Added bindings for animations #26813
  • Fix wrong python type hints for cv::imread #26890
  • Replaced tostring() and fromstring() call in tests and samples for compatibility with modern Python #26932, #26949
  • Replaced ndarray.ptp() with np.ptp() for NumPy 2.0 Compatibility #27133
  • Extended Java bindings generation in DNN #27228, #27260, Features2d #27245
  • Fixed std::vector<Point3f> handling in JS wrappers #27291
  • Added cv::VideoCapture buffered stream constructor into Java #27284, #27421
  • Extended bindings for cv::HoughLinesWithAccumulator #27377
  • Fixed Kotlin 2.0 internal error for unsafe coercions #26354

• Platforms support:

  • Fixed OpenBLAS detection on Linux #27029
  • Fixed compatibility with CMake 4 #27192
  • Fixed Android build with Vulkan support #26765
  • Fixed VSX intrinsics (Power) implementation #26750, #26786
  • Added CMake checks for getauxval and elf_aux_info for POWER architecture #26798
  • Fixed build for openBSD #26803
  • Added initial support for Nvidia Blackwell GPU arch in CUDA #26820
  • Fixed cv namespace dependency in Carotene HAL #26878
  • Fixed _DEBUG/NDEBUG handling across modules #26151
  • Fixed build for QNX 7.0 #26985
  • Switched to statically linked FastCV #26917, #27340 fixed build issues on Android #27345
  • Forced C++17 Standard for CUDA targets when CUDA Toolkit >=12.8 #27112 #27142
  • Fixed build for Windows ARM64EC #27416
  • Fixed filesystem support under Cygwin environment #27484

January, 2025

New Year update for OpenCV 4.x has been released.

• Generic:

  • Internal C API cleanup and back-ports for 5.x.
  • RISC-V/AArch64: disable CPU features detection #25901
  • Support C++20 standard #26590
  • algoHint parameter for some functions to allow potentially faster, but not bit-exact implementation

• Core Module:

  • Added int64 data type support for FileStorage #26434
  • Fixed invalid attribute value handling in FileStorage #25955.
  • Extended LUT for FP16 support #25787.
  • Fixed stdDev tail filling with zeros with HAL in meanStdDev #25789.
  • Set and check allocator pointer for all cv::Mat instances #25979
  • Improved accuracy of Rect::contains #26022
  • Fixed result offset in minMaxIdx with HAL in some cases #26080
  • Replaced C++ operators with wrapper functions on universal intrinsics backends #26109
  • Extended cv::TickMeter #26212
  • Rewrote OpenCL-OpenGL-interop device discovery routine without extensions and with Apple support #26281
  • Marked cv::Mat(Mat&&) as noexcept #25899
  • Multiple Eigen library interop improvements #25751.

• Calib3d module:

  • Multiple chessboard detector improvements #25427, #25807, #25991, #26014
  • Enabled checkerboard detection with a central / corner marker on a black tile #25808
  • Fixed Rodrigues CV_32F and CV_64F type mismatch in projectPoints #25824
  • Added fisheye::distort with non-identity projection matrix #25943
  • SQPnP solver updates #26219
  • Fixed vector access in USAC #26530

• Imgproc Module:

  • Added a new function that approximates the polygon bounding a convex hull with a certain number of sides #25607
  • Added Weighted Hough Transform #21407
  • Fixed bug in contours approximation #25672
  • Fixed bug in divSpectrums #25738
  • Fixed result buffer overflow in intersectConvexConvex_ for non-convex input #25725.
  • Added flag to GaussianBlur for faster but not bit-exact implementation #25792
  • Added flag to cvtColor for faster but not bit-exact implementation #25932
  • Fixed fillPoly drawing over boundaries #26187

• DNN Module:

  • [GSoC] Blockwise quantization support #25644
  • Faster implementation of blobFromImages for cpu nchw output #26127
  • DNN optimization with RISC-V RVV #25883.
  • Added DepthToSpace and SpaceToDepth #25779
  • Yolo v10 support and related samples #25794
  • Parallel implementation of nary elementwise operations #25630
  • Support for Unflatten operation required by Attention layer #25861
  • Erf and GELU layers optimization #25147
  • Activations optimization with v_exp #25881
  • Fixed compilation errors with different OpenVINO versions #25987
  • Fixed matmul crash with CuDNN #26029
  • Fixed CuDNN runtime version check for CuDNN 9+ #25841.
  • Added ONNX TopK #23279
  • Fixed buffer allocation in einsum (fixed random crash on 32-bit platforms) #26059
  • Added Leaky RELU support for TFLite #26132.
  • Switched to run-time dispatcher for Winograd #26155

• Objdetect module:

  • Properly check markers when none are provided #25938.
  • Fixed invalid vector access in QR decoder and encoder #26532

• Highgui module:

  • Added new Highgui backend on top of Framebuffer #25661
  • Fixed HWND_TOP window handling on Windows #25836
  • [GSoC] Added OpenGL support with GTK3 #25822
  • Several OpenGL related fixed on Linux #25865, #22836
  • Fixed leak in cvGetWindowRect_COCOA #26625

• Imgcodecs module:

  • [GSoC] New API for Animations with WEBP, AVIF and Animated PNG support #25608, #25715.
  • [GSoC] Add GIF decode and encode for imgcodecs #25691
  • Added experimental JPEG XL (jxl) codec support #26379
  • Initial RGB layout support in imread and imdecode #25809
  • Fixed imread output type for some cases #25703
  • Fixed file descriptor leak in HDR decoder #25838
  • Fixed corrupted JPEG decoding #25864
  • Improved error handling in image codecs #25975
  • Fixed Bayer2Gray SIMD for U8 #25968
  • Avoid uninitialized value read in AVIF #26084
  • Implemented imencodemulti() #26211
  • Fixed compatibility with different OpenEXR versions #26478.

• VideoIO module:

  • Added VideoCapture constructor for in-memory data stream #25584
  • Fixed bugs in native video encoding on Android #25716.
  • Added BGRA streams support in GStreamer back-end #25602.
  • Updated materials for Orbbec cameras support #25813, #25845, 25907
  • Fixed cv::VideoWriter with FFmpeg timestamps encapsulation #25874
  • Fixed memory leak in Dshow back-end #26045
  • Fixed V4L NV12 color conversion #26081
  • Android native camera feature enhancements #26627
  • AndroidMediaNdkCapture pixel format enhancement #26656
  • Fixed VideoCapture fails to read single image with digits in name #26637
  • Fixed writer setProperty with FFmpeg plugin #26652

• Video module:

  • Fixed VitTrack in the case where crop size grows until out-of-memory when the input is black #25771

• Features2d:

  • Fixed out of bounds access in SIFT #26148

• G-API module:

  • Fixed input buffer read overflow in vectorized G-API convertTo implementation #25780.
  • Extended G-API onnx::Params to pass arbitrary session options #25791
  • Handling I32/I64 data types in G-API ONNX back-end #25817
  • G-API: Introduce level optimization flag for ONNXRT backend #26293

• Optimizations:

  • New FastCV-based HAL for Qualcomm SoCs (-DWITH_FASTCV=ON CMake option) #26556, #26617, 26619, 26621.
  • Added own vectorized version of v_exp #24941, v_log #25781, v_erf #25872, v_sin and v_cos #25892.
  • KleidiCV HAL for ARM updated to KleidiCV 0.3 #26623 and it is turned on by default for Android builds. The supported operations and parameters are documentedhere.
  • Initial version of HAL for RISC-V RVV 1.0 and RISC-V RVV 0.7.1 extensions #26216, #26624, 26675.
  • Used LMUL=2 in the RISC-V Vector (RVV) backend of Universal Intrinsic. #26318
  • More functions in NDSRVP HAL for RISC-V P extension #25786, #26088, #26364.
  • Updated built-in IPP to version 2021.12. Fixed build issues with old and new Intel IPP layouts #25773

• Platforms:

  • HWAsan support on Android #25746
  • Several CUDA fixes for old GPUs without FP16 support #25880
  • Added getStdAllocator() to cv::cuda::GpuMat #25978
  • Updated NPP calls to use the new NppStreamContext API if available #26245
  • More convenient GpuMatND constructor #26472
  • Added run-time GPU check to haveCUDA #26164.
  • Add support for QNX #25832

• OpenCV.js

  • Extended API white-list and added more tests #26387
  • Split white-list per module. Added opportunity to cover opencv_contrib modules #25986
  • Fix incorrect string format in js build script #26374
  • Emscripten build fixes with SIMD intrinsics #26537
  • Added more public types for USAC support #26638
  • Rename Mat::clone binding because it is used in Emscripten #26643
  • Fixed C preprocessor stringification #26644
  • Fix enum generation issues #26147
  • Multiple test improvements #25732, #25757, #25961

git shortlog --no-merges -ns 4.10.0..4.11.0
    57	Alexander Smorkalov
    34	Maksim Shabunin
    25	Vincent Rabaud
    16	Rostislav Vasilikhin
    15	Suleyman TURKMEN
    13	Kumataro
    12	Letu Ren
    10	Yuantao Feng
    9	Maxim Smolskiy
     8	Dmitry Kurtaev
     8	Alexander Lyulkov
     7	Liutong HAN
     7	kallaballa
     5	Alexander Alekhin
     5	Wanli
     4	FantasqueX
     4	cudawarped
     3	Abduragim Shtanchaev
     3	Rüdiger Ihle
     2	Amir Hassan
     2	Hao Chen
     2	Junyan721113
     2	KangJialiang
     2	Manolis Lourakis
     2	Maxim Milashchenko
     2	Pierre Chatelier
     2	chacha21
     2	eplankin
     2	ericmariasis
     2	richard28039
     2	zihaomu
     1	Alberto Zurini
     1	Alessandro de Oliveira Faria (A.K.A.CABELO)
     1	Alexander Panov
     1	Aliaksei Urbanski
     1	Andy-SAKN
     1	Arvind Iyer
     1	Aven
     1	Billy Robert O'Neal III
     1	CSBVision
     1	Christine Poerschke
     1	Corentin Noël
     1	Dan Mašek
     1	Daniele Affinita
     1	Dietmar Schabus
     1	Dmitry Yurov
     1	HAN Liutong
     1	Hardik Kamboj
     1	Ivan Afanasyev
     1	James Choi
     1	Javier Macias Sola
     1	Karsten Wiese
     1	Kimberly N. McGuire
     1	Masahiro Ogawa
     1	Miguel Daipré
     1	Mikhail Khachayants
     1	Mironov Arseny
     1	MurtazaSaherwala
     1	Neko Asakura
     1	Nishanth
     1	Octopus136
     1	Onur Can KARAMAN
     1	Oちゃん
     1	Patrick Keane
     1	Philip Lamb
     1	Richard Barnes
     1	Robert Lexmann
     1	Robert Mitchell
     1	Scott
     1	Septimiu Neaga
     1	ShengYu
     1	Simon Kämpe
     1	Skreg
     1	Sourav Kumar
     1	Super
     1	TFiFiE
     1	Thirumalai Nagalingam
     1	Ujjayant Kadian
     1	Vadim Levin
     1	WU Jia
     1	WangWeiLin-MV
     1	Yang Chao
     1	Yannis Guyon
     1	Yoshiki Obinata
     1	Zach Lowry
     1	_Ayaka
     1	adsha-quic
     1	alex-urm
     1	anandkaranubc
     1	cDc
     1	catree
     1	fengyuentau
     1	gblikas
     1	george
     1	inayd
     1	j3knk
     1	kaingwade
     1	kozinove
     1	lamiayous
     1	llh721113
     1	pasbi
     1	penghuiho
     1	quantizs
     1	quic-apreetam
     1	quic-xuezha
     1	shengyu
     1	shyama7004
     1	stepkamipt
     1	tingboliao
     1	xkszltl
     1	xndcn
     1	武士风度的牛

December, 2024

The alpha release for the new OpenCV generation. The release is designed as technology preview and not ready for production usage yet.

The following notation is used to mark various items below:

• 5+4.x: the feature is not quite new, it's already put into the latest 4.x, but it's important enough to mention it here as well.
• 5.0: the item will be finished in 5.0 release; currently it's planned or in progress.
• 5.x: the item will be put in some form into 5.0 release, but will mostly be finished in subsequent 5.x releases.

• 5+4.x: Recent OpenCV 4.x (since OpenCV 4.5.0) and also OpenCV 5.0 are now distributed under Apache 2 license. We used BSD license before, but switched to Apache 2, because the latter is more refined and provides somewhat better protection (not ideal though) from patent-related issues.
• OpenCV now requires C++ 17 as a minimum C++ standard. It is built by default with C++ 17, but we also plan to make it compatible with C++ 20 and C++ 23.
• Python 2 support is removed. OpenCV now requires Python 3 (3.6+) and we only build Python 3 bindings.

• C API has been removed. We still use some CV_ macros, such as CV_8U, but all C functions (like cvCreateMat() or cvFindContours()) and structures (such as CvMat) have been removed. Goodbye, good old OpenCV 1.x API, RIP :)
• OpenVX support has been removed. If some vendors provide OpenVX kernels and want to use them for OpenCV acceleration, they can create a custom “non-CPU” HAL for that, see below.
• Graph API (G-API) module has been moved to opencv_contrib.
• Classic ML module has been moved to opencv_contrib. If you use Python, scikit-learn is a much better alternative.
• Features2D module has been renamed to Features. Scope of the renamed module has been extended to process feature vectors produced by the modern deep nets. Several obsolete feature detectors/descriptors have been moved to opencv_contrib. SIFT, ORB, FAST, GoodFeaturesToTrack, MSER are still available, though.
• 5.0: FLANN as a separate module will be gone. Annoy-based ANN (approximate nearest neighbor) search algorithm, which has been already added to Features module, will replace it.
• Cleaned objdetect module: Haar-based and HOG-based detectors have been moved to opencv_contrib, xobjdetect module. There are modern deep learning-based object/face detectors that are both faster and more accurate.
• Calib3d module has been split into 3 modules:
  • 3d — basic 3D geometry and 3D vision functionality
  • calib — camera calibration
  • stereo — algorithms for depth map estimation via stereo correspondence.
• Many obsolete samples (~50% of C++ samples and 5% of Python samples) have been removed. Many of the remaining samples have been revised.

• OpenCV now supports the extended set of data types:
  • uint8_t: CV_8U, int8_t: CV_8S, uint16_t: CV_16U, int16_t: CV_16S, int32_t: CV_32S, float: CV_32F, double: CV_64F, hfloat (a.k.a. half or float16_t or __fp16): CV_16F.
  • (new in OpenCV 5): bfloat (a.k.a. bfloat16_t): CV_16BF, uint32_t: CV_32U, uint64_t: CV_64U, int64_t: CV_64S, bool: CV_Bool.
  • bool type takes 1 byte per value, not 1 bit. Any non-zero byte is considered true, zero byte means false. cv::Mat of type bool (CV_Bool) can now be used as a mask for all functions where we used cv::Mat of uchar/uint8_t or schar/int8_t before.
  • Operations on hfloat and bfloat are always available, even on hardware that does not support those types natively. If necessary, internally we use efficient inline functions for scalar and vector float<=>hfloat/bfloat conversion. Support for the new types has been added to cv::Mat, cv::UMat, InputArray/OutputArray, key modules (core, dnn, (5.x) imgproc etc.), FileStorage, various programming language bindings etc.
• OpenCV now supports arrays of lower than 2 dimensionality, i.e. 1D (vectors) and 0D (scalars). std::vector<T> wrapped into Mat or InputArray/OutputArray is now a real 1D array, not 2D Nx1 or 1xN array (as in OpenCV 4.x). For 1D arrays Mat::dims == Mat::rows == 1, Mat::cols == Mat::total() == <number_of_elements>. For 0D arrays Mat::dims == 0, Mat::rows == Mat::cols == Mat::total() == 1. The safe method to distinguish between an empty matrix and a scalar is to use <some_mat>.empty(), which is basically equivalent to (but is slightly faster than) <some_mat>.total() == 0.
• Lapack is now always available within OpenCV. In particular, it’s used for more efficient SVD and eigenvalue/eigenvector decomposition and the USAC framework. When there is no external Lapack library installed in the system, OpenCV builds and uses the internal small subset of Lapack.
• 5.0, 5.x: Further refactoring and improvements are expected before and after 5.0 release: https://github.com/opencv/opencv/issues/25011.

• Accelerated image warping functions: warpAffine, warpPerspective, remap. Acceleration factor varies from 10% to more than 300%, depending on the platform, image size, type and operation flags. More optimizations for those functions, as well as other important image processing functions are expected in 5.0 gold.
• Text rendering now uses STB-based truetype engine and the embedded variable truetype font. Users can also load and use custom fonts. Consequently, many Unicode symbols are now supported, with a few caveats though:
  1. some scripts, such as Arabic or Devanagari, are not rendered properly, we need to add Harfbuzz for that
  2. some compound Unicode symbols are not rendered properly for the same reason
  3. color emoji are not supported, as STB is a simple black-n-white engine.
• 5.0, 5.x: Further refactoring and improvements are expected before and after 5.0 release: https://github.com/opencv/opencv/issues/25012

• 5+4.x: Many new entries have been added to OpenCV HAL, allowing various vendors to provide custom acceleration to OpenCV functions. In 5.0 gold some more new HAL entries will be added.
• 5+4.x: In some cases vendors provide super-fast implementations that are unfortunately not 100% compatible with OpenCV. To let user to choose between speed and accuracy in some critical places, OpenCV introduced the optional AlgorithmHint hint parameter to several functions, which is set to ALGO_HINT_DEFAULT by default (which is equivalent to ALGO_HINT_ACCURATE unless user compiled OpenCV with ALGO_HINT_DEFAULT=ALGO_HINT_FAST (-DOPENCV_ALGO_HINT_DEFAULT=ALGO_HINT_APPROX option in cmake).
• 5+4.x: Universal intrinsics for math functions (v_exp, v_log, v_erf, v_sincos) have been added to accelerate deep learning inference, image processing and other algorithms.
• FP16 (i.e. half-precision) universal intrinsics have been added into ARMv8 NEON and RISC-V RVV backends. In both cases FP16 SIMD arithmetics is not available by default, the code needs to be built with special options. The corresponding parts of OpenCV use our runtime dispatching mechanism, which chooses the best-suited flavors of kernels depending on the actual client hardware.
• 5.0, 5.x: UMat will be extended to be able to store any CPU or non-CPU array/tensor. That is, from an OpenCL-only solution (T-API) we now migrate to a universal heterogeneous API (we might call it U-API or “non-CPU HAL”) and UMat will be the main data container for it. Many other improvements are expected as well: https://github.com/opencv/opencv/issues/25019, https://github.com/opencv/opencv/issues/25025

• The new engine has been introduced that now co-exists with the old engine. The new engine provides better support for dynamic shapes and other modern ONNX features. By 5.0 gold it will also provide much better coverage of ONNX specification than the old engine; now the coverage is comparable.
• Added parameter int engine = ENGINE_AUTO to cv::dnn::readNet() to control which engine is used to load and further run the model. By default, we try the new engine first and if it fails to load the model, fall back to the old engine. We can also force the new engine or the old engine. Because of the different internal representation the engine cannot be switched after the model is loaded.
• Parsers for ONNX, Caffe, TF and TFLite formats have been updated to support the engine selection.
• 5.0, 5.x: currently, the new engine only supports the default backend and CPU target. It’s planned to enable more backends and targets by 5.0 gold and even more backends in subsequent 5.x releases #26198.

• 5+4.x: Significantly more efficient USAC framework (especially in the case of noisy datasets) is now used for RANSAC-based algorithms to estimate homography, essential matrix, solve PnP problem etc. See https://docs.opencv.org/5.x/de/d3e/tutorial_usac.html
• New Levenberg–Marquardt algorithm implementation. It's faster and more accurate.
• Efficient multi-camera calibration framework has been added to calib module. It includes smart initialization and USAC-based optimization pipeline to compute relative positions of all simultaneously calibrated cameras w.r.t. the 1st camera. Various camera configurations are supported (all pinhole cameras, all fisheye cameras and the mixed case). See tutorial for more details.
• Initiated work on basic mesh processing and point cloud processing algorithms: TSDF, ICP, etc.
• Added importers and exporters for some popular point cloud formats, such as .ply and .obj.

• Revised the following deep learning samples: classification, segmentation, object detection, edge detection, tracking, person reid. All the samples use a collection of models that can be conveniently downloaded with cd <opencv>/samples/dnn && python3 download_models.py.
• Added experimental samples for LLM (GPT2) and diffusion (LDM) models.

TBD