mirrors/opencv.opencv

Fork 0

mirror of https://github.com/opencv/opencv.git synced 2026-09-06 04:25:51 -05:00

Table of Contents

2026 OpenCV Technical Committee
Template

2026-mm-dd

Minutes

2026 OpenCV Technical Committee

Meeting_notes

Template

2026-mm-dd

Minutes

2026-08-26

Abhishek and BigVision Team:
- Abhishek:
  - current ONNX coverage is 76%
  - raised PR #29844, #29834, #29832, #23830 (batch resize, simd support for fp8, fp8 support for DNN, AVX-VNNI).
  - merged #29594, #29360 (bitcast layer, matmul, dft, cast/castlike layers)
  - currently working on non-CPU HAL.
- Savya:
  - merged #29658 (introduction of cuDNNJIT)
  - raised #29843 (added camera device enumeration in Video I/O)
  - working on replacing NPP kernels with native ones: flip, transpose, threshold, box filter.
- Ayush:
  - merged: PR #29767, #29786, #29835, #29049: python int32 fix, extended data types in elemwise, gather and scatter, PPL auto_partitioner, IPP bilateral function moved to HAL.
  - raised: PR #29833, #29841: Windows ARM64, promote NAryElemwise to common type.
  - adding batch support to DNN
  - locally testing Windows ARM-related PRs
- Varun:
  - merged: PR #29806, #29782.
  - raised: PR #29675: refactored tokenizer
  - inter_area for warping.
  - collected test execution time, probably we can reduce overall test execution time by prunning it.
- Om Rope:
  - working on AWS COOL:
    - COOL for AWS Graviton 5
      - had to temporarily disable LAPACK. investigating the issue
  - adding cv::linearAssignment to core. Alexander suggested to look at labjv.
Vincent:
- filed 2 bug reports:
  1. Array of distortion parameters of size 4 is not processed correctly.
  2. Issue with licenses.
- the new font rendering engine takes significant amount of space.
Gary:
- started to work at Stanford.
Satya:
- contract with AMD is being finalized.
- conversation with Intel goes on, but not quite actively.
Gursimar:
- Windows ARM binaries: need to test on Qualcomm devices and prepare release (5.1 alpha?)
Phil:
- OpenCV AI competition started 2 days ago.
  - 117 participants already
  - Received 183 submissions (160 of them are valid)!
- 'Fuse' conference starts in SF soon. They want to do extra advertisement via OpenCV live.
- Communications with Qualcomm go smoothly, the project has finally started.
- Talking to several existing OpenCV members/sponsors, expect that membership in most or all cases will be extended for another year.
- And also talking to several new potential OpenCV membmers.
- OpenCV live is booked until 2027 Jan :)
Alexander S (opencv.ai):
- working on OpenCV Python.
  - ported CI actions completely to github actions
  - because of it, OpenCV Python on macos is now tested on macos 15+, we don't support older versions any longer
- added another server in SusTech, but the older server does not well, seems to be a problem with disk
- significant progress with optimization PRs
  - merged many PRs from Qualcomm and AMD developers
  - fixed Concurrency issue on Windows ARM.
- merged many DNN improvements from BigVision (Abhishek & co).
  - 2 big PRs, especially on CUDA, is still awaiting review
- planned to do the next 4.x to 5.x merge, and it will be one of the latest 4.x -> 5.x merges, because it becomes increasingly more difficult to do those merges.

2026-08-26

Abhishek and BigVision Team:
- Abhishek:
  - #29785 (Add image decoder ONNX layer)
  - #29642 (dynamic kv-cache support)
  - #29624 linear and flex attention
  - #29595 (replace global finalizeLayers flag with per-layer re-initialization)
  - #29784 (alpha composite)
  - #29783 (added gridsample bicubic, dropout support)
  - #29594 (added bitcast & extended matmul & dft layers)
  - #29360 (support ONNX Cast/CastLike for FP8/FP4/INT4/UINT4/E8M0 dtypes)
  - working on batch resize support.
- Savya (BigVision intern):
  - PR #29732 (alexnet memory consumption reduction by revising onnx importer)
  - PR #29764 (documentation extension)
  - PR #29656, #29658 refactoring backend-agnostic fusion & cuDNNJIT.
- Ayush:
  - merged PR #29743, #29752, #29753, #29410: calib, features, geometry tests; IPP integration of biliteral filter
  - raised PR #29767, #29786: python fix int32, extended data support in elementwise, gather & scatter layer.
  - built experimental OpenCV 5.x Python experimental binaries for Windows on ARM.
  - working on Windows ARM-related PRs, fixing NARyEltwise type mismatch
- Varun:
  - #29782, extra #1406: Data type support extension in DNN
  - #29806 fxies for Windows on non-ASCII path handling in ONNX runtime.
  - #29675: refactored tokenizer
- Om Rope:
  - working on AWS-COOL:
    - OpenCV 5 for Graviton 5 is ready for release.
      - custom AMI and docker containers have been prepared
      - valided across 102 ops
  - working on multi-object tracking.
Vincent:
- continue to play with OpenCV 5 internally
Phil:
- preparing for the next conferences
- OpenCV podcast is booked till 2026 Nov; this week there will be guests from Intrinsic.ai company (3d perception topics: https://www.intrinsic.ai/publications/3pt-cvpr2026).
- multiple new press-releases packages are being prepared:
  - on AWS competition
  - on Qualcomm partnership
  - on AMD partnership
  - on collaboration with roboflow (https://summit.roboflow.com/) and other companies as well
- people could donate to OpenCV directly from Youtube.
- OpenCV patreon program does not go well yet, but still big thanks to those supporters
Gary:
- opencv.org home page shall probably be reorganized.
- doing some interesting stuff at Stanford, will see how this can be connected with OpenCV.
  - get better 3D perception results by asking models to explain what they see from 3D geometry perspective.
  - research on next-gen VLMs
  - thinking of cheaper arms and other essential robot components
Alexander S (opencv.ai):
- Done:
  - Alexnet and other models now fit memory better; the new patch reduces the memory consumption by 2x.
  - ONNX coverage is further improved.
  - dynamic CUDA part is merged.
  - working on adding support for more cameras (Basler etc.). Usually vendors use proprietary SDKs for the cameras. There is open-source Aravis framework to cover some of those hardware, we are trying to make wrapper for Aravis. Aravis is LGPL, so the plugin needs to be a shared library.
  - continue to improve CI.
    - Xeon servers that we use in SusTech are ~4x slower than the newer servers in BigVision
    - suspected that we have troubles with ccache, but it appears that we have 90% hit rate with ccache.
      - but we can further improve reduce build type: cmake part is still slow, and sometimes we miss ccache - need to improve hit rate.
      - 32-bit linux is also slow.
      - classic DNN engine can be excluded from OpenCV 5.x tests.
    - going to add Alpine linux to CI.

2026-08-19

Abhishek and BigVision Team:
- Abhishek:
  - reviewed several patches from Qualcomm, suggested to use universal intrinsics more heavily instead of custom branches
  - merged PR #29674 (much faster LSTM)
  - dynamic kv-cache PR #29642
  - per layer re-initialization PR #29595
  - DFT and bitcast PR (#29594)
  - xfeat #29361
  - disk + lightglue PR #29392
  - SLAM & loop closure : #29650
- Savya:
  - PR #29732 fixed alexnet
  - PR #29729 added gaussian splatting
  - PR #29656 adding backend agnostic fusion
  - PR #29658 CUDNN JIT support
  - working on hybrid CPU/GPU fallback
- Ayush:
  - Merged #29666, #29702, #29719: matmulnbits, DNN tests, re-enable & triage DISABLED tests.
  - enabled CI for Windows ARM
  - testing locally OpenCV + Windows ARM patches submitted by Qualcomm team
- Varun:
  - working on test cleanup: PR #29790 imgproc, #29741 objdetect, #29742 ovideo, photo, ptcloud
  - extending DNN to support more data times (Range, Hardmax, MaxUnpool, CumSum/CumProd, MaxPool)
  - Sphinx OpenCV docs: search query optimization
- Om Rope:
  - AWS-COOL: continue work on optimization work for Graviton 5
    - benchmark of 102 demonstrates 23% perf gain vs vanilla OpenCV 5
Phil (opencv.org):
- New OpenCV competition is here now with some big sponsors: opencv26.devpost.com
- 'fused' conference in 2026 Nov; will attend it.
- British Machines Vision renewed their sponsorship
- OpenCV live streams are getting more popular, have invited guests queue for the next 3 months
- talking about next OpenCV competitions.
- working on the next OpenCV fund raising programs.
  - OpenCV 5 'cookbook' will probably be the subject of one of such programs
- will participate in OSCCA conference next year.
Alexander S (opencv.ai):
- switched to 5.x as the defaul OpenCV branch
- busy with reviewing many PRs
- there is issue with CI, related to some long-term contributors representing different organizations - currently their PRs still need manual approval.
- system Lapack sometimes break calibration pipelines.
- ffmpeg binaries for macos include some GPL components
- cascade classifiers and g-api lost bindings because those things have been moved to opencv-contrib. Now restoring it.
- macos builds need more power; probably will use cloud CI servers provided by github.
- large patch by Mu Zihao: lot's of DNN accelerations for AMD GPUs
- experimenting with CUDNN JIT, have some troubles with building it

2026-08-12

Abhishek and BigVision Team:
- Varun:
  - working on CI:
    - PR #29657: python conformance tests
    - PR #29630: Windows CI
  - working on non-CPU HAL, e.g. adding support yolo v8/v5 and yolox. Pipelines completely run on GPU, without falling back to CPU.
  - AlexNet takes too much memory.
- Savya:
  - PR #29656 added backend-agnostic fusion
  - PR #29658: CUDNN JIT: 9.3x speedup on Resnet 50 (RTX 6000 vs Intel Xeon).
  - Hybrid CPU/GPU fallback for ONNX runtime graphs
  - Gaussian splatting is in progress
- Jaivardhan:
  - PR #29675: adding support for more tokenizers, all kinds of BPE, SentencePiece, WordPiece, Unigram
  - added samples for BERT (wordPiece), ALBERT (unigram)
- Ayush:
  - PR #29666 Added MatMulNBits layer and extended ONNX coverage
  - PR #29702: re-enable & triage a few 'DISABLED_' tests
  - also submitted the corresponding opencv_extra PRs
  - elemntwise, slice2, gather_elements, gatherND, scatterND, scatter layers.
  - working on 'Windows on ARM' build of OpenCV.
    - Alexander S take on OpenCV for 'Windows on ARM':
      - there are still some stability issues (but according to Ayush we don't have regressions now)
      - there are still some performance issues (can be postponed for the first release)
      - suggest to use old approved variant with 'opencv_megapack.exe' instead of suggested vcpkg or something like that.
      - suggest to use regular Microsoft compiler instead of clang, even though MSVC is known to have problems with building reliable code for ARM.
- Om Rope:
  - working on AWS-COOL
    - started optimization for Graviton 5
    - revised benchmarking (increased number of compared kernels from 78 to 102)
      - included more popular operations and removed some obsolete stuff
    - there is already 21.4% gain of COOL vs regular OpenCV 5.0
    - next steps: continue optimization work
Phil (opencv.org):
- BMV (British Machines Vision) is going to renew their participation in OpenCV foundation as a sponsor/partner.
- Will announce AWS-sponsored AI competition very soon (maybe tomorrow)
  - up to 50 teams could participate (the number is limited to make sure OpenCV org can accurately review all the submissions)
- Going to participate in EVS (scheduled to 2027 Feb), going to have a booth there + speakers
- As well as take part in CVPR 2027
- OpenCV 5 cookbook, a few hundreds pages. Probably Gary can write a preface and maybe a couple of chapters for it
- OpenCV light. Still many requests and this is something we need to think about.
  - it can be commerical tool to extract only the needed (customer-specific) subset of OpenCV.
- OpenCV live passed 180K subscribers.
Alexander S (opencv.ai):
- Alexaner was given a free Claude Code subscription as active open-source contributor, and started to use it actively.
- Merged ~20 PRs.
  - Some of the patches ONNX coverage further.
  - Some PRs are not merged yet, need some discussion since they change protocol how OpenCV DNN operates
  - GPU support is being constantly improved.
- After some initial cleanup going to work on some new features to utilize Claude C more intensively.
- Documentation probably needs some further polishing and the documentation distribution and representation should be improved (search in particular).

2026-07-29

Abhishek and BigVision Team:
- Abhishek:
  - DNN engine classic removal is DONE and merged finally.
  - ONNX coverage is 72.9% now
    - Several PRs are pending:
      - Linear and Flex attention,
      - global finalizeLayers flags with per-layer re-initialization
      - added bitcast layer & extended MatMul and DFT layers
  - PRs Merged:
    - PR #29341 [Remove ENGINE CLASSIC, switching to ENGINE NEW as default engine]
    - PR #29451 [Extending CUDA support in UMat]
    - PR #29524 [Fix: 0-D / empty-Mat Python binding]
    - PR #29560 [Updated ONNX conformance list]
    - PR #29577 [Add Scan layer to the new engine]
    - PR #29579 [Added Cumprod and Causalconv layers in new dnn engine]
  - PRs Finalized:
    - PR #29386 [CUDA support]
  - GSoC:
    - Xfeat feature extractor: PR #29361
    - Disk+lightglue: #29392
- Agrim:
  - Modular SLAM:
    - vidual odometry is done, g2o and loop closure are not done yet.
- Savya:
  - OpenCV GPU-accelerated imgproc functionality is migrated to the main repo:
  - PR #29334
  - working on adding GPU-acclerated DNN operations and layer fusion engine
- Ayush:
  - Working in Win @ ARM64 CI:
    - MinGW, assembly
  - extending DNN to support more types
- Varun:
  - working on extending CI
    - regresssion fix #29581, #29369, #29426
    - Yolo 26nn #1357 (extra)
    - PR #29630
  - implementing purely GPU pipelines that do everything in GPU memory
- Om Rope - work on AWS:
  - COOL Docker images for OpenCV 5 are officially released.
  - support Python 3.10, 3.11, 3.12
  - fixed various security issues
  - CI/CD pipeline is built for automatic testing:
  - end-to-end automated testing pipeline for Graviton 4 hardware.
    - plan to extend it to Graviton 5.
Jiale Yang (opencv China):
- working on RISC-V RVV optimization.
- the latest version demonstrated not very good performance on K1
- working on fisheye camera patches
Phil (opencv.org)
- working on upcoming competitions
  - Satya has attended AMD conference in SF last week, had quite productive meetings with AMD folks.

2026-07-22

Vincent:
- Raised PR to silent MSAN warnings.
Abhishek and BigVision Team:
- Abhishek:
  - Working on expanding ONNX layer coverage including linear_attention and cumprod.
  - Finalized PRs for CUDA support and Engine Classic Removal.
- GSoC:
  - Finalized PRs for Xfeat and Lightglue matcher integration.
  - Completed Visual Odometry SLAM cleanup.
- Savya (Bigvision-Intern):
  - Adding CUDA support to the new DNN engine focusing on UMat memory integration to optimize device data transfer speeds.
  - Implementing CuDNNJIT compilation and custom CUDA kernels for unsupported deep learning layers.
- Ayush:
  - Extended the TFLite importer to support more models on the new engine and added TFLite test data.
  - Added Ultralytics YOLO models in TFLite format via a merged Hugging Face PR.
  - Debugging Windows ARM CI multi-threading build failures locally on a Qualcomm device.
- Varun:
  - Extended GPU HAL dispatch for pre-processing and post-processing functions.
  - Achieved a 2.8x total speedup on the ResNet50 pipeline, inclusive of device transfers.
  - Waiting on UMat CUDA support to eliminate transfer timings and retest.
- Om Rope:
  - Designed the automated testing pipeline and Dockerfile for Graviton 4, currently testing locally.
  - Cleared CVE security checks for Python 3.10-3.12 and pushing containers to ECR (Releasing Docker version of COOL).
Gary:
- Just returned from vacation and has started at Stanford.
Phil:
- Attending the AMD conference in San Francisco with Satya.
- Working on the AWS Devpost competition.
Zihao:
- Currently working at AMD and offered to assist with testing the CI on physical AMD GPU machines.
- Testing Imgproc functions on AMD CPU and GPU and getting upto 40x compute speed up.
Jiale Yang:
- Working on RVV support but currently facing network issues with the K1 board at NUS.
- Will resume testing upon returning to China in August.

2026-07-08

Vadim:
- working on dynamic CUDA support in OpenCV
Abhishek Gola & the team (BigVision):
- Abhishek:
  - raised PR at onnx.ai project (https://github.com/onnx/backend-scoreboard/pull/131) to fix coverage: it's now 72.1%
  - PR #29524. fix: 0-D/empty-Mat Python bindings
  - other merged PRs:
    - PR #29288: new documentation fixes
    - PR #29369: FP8 support
    - PR #29516: added missing MLAS files for IBM power
  - PR submitted: #29386 (CUDA support), #29341 (remove classic DNN engine)
  - GSoC: Xfeat feature extractor, PR is raised #29361, Disk + lightglue PR is almost ready.
- Agrim (Bigvision intern):
  - Visual odometry #29331
  - Next goal is G2o and loop closure
- Akansha (Bigvision intern):
  - moved a bunch of IPP functions to IPP HAL.
- Kirti (Bigvision intern):
  - PR #29485 OpenGL wrapper for HighGUI
  - after it's merged, will add 2 more patches: 3D visualization samples, ptcloud processing functions.
- Ayush (Bigvision intern):
  - PR #29434 (canny_ipp_extract)
  - PR #29515 (yolov3_replace)
  - testing Windows ARM:
    - 3 PR's have been reviewed.
- Savya (Bigvision intern):
  - PR #29474 (TF importer removal from 5.x)
  - working on: cuDNN JIT support
  - integration combines unidifed UMat.
- Varun:
  - CI fixed: #29499 merged: videoio and core module tests fail on Windows 11.
  - Ubuntu 26 CI failures: opencv_test_calib 20 min timeout.
    - AVX-2 vs AVX-512 frequency throttling.
- Om Rope:
  - Working on COOL, Graviton 3. 21.76% speedup over vanilla OpenCV 5.0.
  - applied some security patches
  - prepared final performance spreadsheet 4.13 vs 5.0. Will be reviewed internally, then published.
Vincent:
- still bumping OpenCV internally.
- a few calibration things need more investigation, looking at core and dnn now.
Phil (opencv.org):
- OpenCV AWS-sponsored competition is ahead.
- Talking to a company who organize 'position & navigation' conference. Satya will give a talk there.
- Negotiations with Qualcomm go on.
- Negotiations with AMD are going well.
Alexander S (opencv.ai):
- 4.14 will be released this week
  - fixed ffmpeg bindings
  - have sporadic failures with OpenVINO backend, difficult to reproduce.
- 5.x: BigVision folks continue to move IPP calls to dedicated HAL.
  - Merged several patches from AMD team with accelerations
  - Merged OpenCL kernel for SIFT
  - Merged FP8 support
  - Measured ONNX coverage via standard ONNX compliance test suite: initially got 38% coverage, but with a few fixes in the Python script that measures the coverage we reached 72.1+% coverage.
- CI: now have issues with downloading DNN extra models. Going to fix it this week.

2026-07-08

New members to OpenCV team (not all of them are actually new):
- Mu Zihao: was a part of OpenCV team a few years ago, now working for AMD, but can contribute to OpenCV
- Yang Jiale: Shiqi's student from SusTech, working on (RISC-V + RVV) support in OpenCV.
Vadim:
- submitted arithmetic operations patch to 5.x
- working on dynamic CUDA support in OpenCV
Abhishek Gola & the team (BigVision):
- Detected several performance regressions in OpenCV 5.0 vs OpenCV 4.x. Nothing dramatic, but still worths looking at: https://docs.google.com/spreadsheets/d/1rQ4jQhaxLpi9dcLS7sZYP6_BL5Qe_5kjv4399H7Av60/edit?usp=sharing
- PR #29451 (CUDA support)
- PR raised: #29288 (new docs), #29369 (FP8), #29341 (engine removal)
- refactoring CUDA PR to use GpuMat instead of BackendWrapper.
- MLAS support (#29465)
- GSoC: feature extractor & matcher, almost complete (#29361, #29392)
- Agrim (BigVision intern):
  - Visual odometry PR #29331. Improved modularity.
  - Next goal G2o and loop closure.
- Akansha (BigVision intern):
  - Fp8 patch is prepeared, ready for review
  - multiple PRs on IPP to HAL transition.
- Kirti (BigVision intern):
  - OpenCL integration and 3D visualization (resuming work in #20371 by GsoC student several years ago)
  - some complex point cloud processing functionality will be brought to ptcloud module.
- Savya (BigVision intern):
  - PR #12 (huggingface opencv):
    - converted TF models to Onnx.
  - also removed TF models from opencv, opencv_contrib, opencv_extra. Converted to onnx.
  - working on cuDNN JIT for opencv dnn: https://pypi.org/project/nvidia-cudnn-jit-cu12/
- Varun:
  - working on OpenCV CI. score-driven CI priority layer for 5.x.
  - working on Windows CI issue.
  - non-CPU HAL: #29334 (resize, cvtColor run on CUDA automatically via standard cv::resize)
    - convertTo(), subtract() and divide() are being added as well.
- Om Rope:
  - working on AWS (AWS-COOL). Discovered some perf regressions and now fixing them.
  - audit for 3rd-party libs used by OpenCV: ffmpeg, pthreads, KleidiCV.
Shiqi Yu (OpenCV China):
- no major updates
Alexander S (opencv.ai):
- Finally OpenCV 5.0 Python package has been released; everybody is welcome to try it.
- Working on the upcoming 4.14
- Reviewing many, many PRs related to optimization.
  - part of the PRs are about moving IPP outside of the main tree to dedicated HAL; we are not removing IPP (!), we just make IPP integration more modular.
  - we will gladly accept patches both from Intel and AMD, but none of the patches can significantly degrade performance on the 'other side of the moon'.
- Also, reviewing many patches on DNN.
- Preparing patch for the regular 4.x=>5.x merge.
- There is a first bunch of patches to extend UMat with CUDA, HIP and Apple Metal support. We are looking at them, but none of them is generic enough.

2026-07-01

Vadim:
- continue working on extended arithmetic operations in opencv 5.x
Abhishek Gola & the team (BigVision):
- PR 29333 merged (extended attention)
- refactored 29286 to use cleaner naming conventions.
- reviewed Visual Odometry PR (SLAM part 1)
- published updated documentation.
- working on adding CUDA support in UMat.
- will work on adding more CI configurations.
  - probably can use workstations
  - going to reactivate nightly builds of opencv docs.
- GSoC: both PRs are ready, Xfeat and Disk+lightglue.
- Akansha (Bigvision intern):
  - Integrated OpenVINO into the new engine: PR is not submitted yet.
  - Fixed engine classic removal failures in OpenCV contrib.
  - Fixed attention layer issue (PR #29333)
  - FP8 patch is almost ready, FileStorage part is still missing.
- Kirti (Bigvision intern):
  - Big documentation cleanup PR #29263
- Ayush (Bigvision intern):
  - many IPP-related PRs (enable more IPP functions): PR #29397, #29389, #29409, #29410, #29414, #29420.
- Savya (Bigvision intern):
  - PR #29386: added ONNX dynamic quantization feature.
  - now working on native CUDA execution path to DNN.
- Varun:
  - fixed a few CI issues: yolo26n.onnx.
  - added --download_large_model flag to download big models that are not downloaded by default.
  - added score-driven CI priority level to prioritize some light-weight higher-priority jobs.
- Om Rope:
  - working on AWS-COOL.
    - COOL v3 with OpenCV 5.0 binaries is now deployed in AWS. It's 20.56% faster than vanilla OpenCV 5.
    - security threat fixed (CVE-2026-32746).
Shiqi Yu (OpenCV China):
- several students demonstrate very good results; encouraged them to continue work on OpenCV.
- Zihao also contacted Shiqi, probably, he can help with OpenCV as well.
Vincent:
- still testing OpenCV 5 internally, especially calib, geometry and imgproc. The transition appears to be smoother than it seemed to be, which is great.
Phil (opencv.org):
- Continue to promote OpenCV 5. Live streams got a lot of attention.
- AWS-sponsored competition will be launched soon.
- Big OpenCV university courses
- Still talking to Qualcomm about another competition.
- Patreon page is bringing some money (not much, but still).
Gary:
- Just got back from NY; gave a talk at open-source event there (there have been also Yann LeCun and other big folks).
- will be travelling in Europe in the nearest few weeks.
Alexander S (opencv.ai):
- still working on OpenCV 5.0 Python release.
  - most of the issues are fixed, package is being prepared.
  - there was idea to rename opencv-python to opencv, but for now it was postponed till 5.1.
  - later on it's also planned to split solid opencv-python package into parts: opencv, opencv-contrib etc. There will be separate python packages for various platform-specific parts, e.g. support for specific cameras, acceleration for CUDA etc.
- also fixed Android package; rebuilt OpenCV with the latest NDK.
- working on various optimizations, PRs are coming from various people from opencv community.
  - continue to move IPP-related stuff out of the modules to dedicated IPP HAL.
  - many RISC-V+RVV patches have been contributed to accelerate DNN and other functionality.
- will have meeting with IPP guys this Friday.
- got e-mail from Mu Zihao, we will resume our collaboration.
- started working with students in the local lab in NiNo.

2026-06-24

Vadim:
- working on extended arithmetic operations (broadcasting, expressions etc.) in opencv 5.x
Abhishek Gola & the team (BigVision):
- PR 'fp8 support' in progress
- PR 'cast/castlike support for fp8/fp4/int4' raised
- PR 'remove classic DNN engine' is raised
- PR 'disabled MLAS on Windows' merged
- PR 'extended support for Resize and Split' merged
- PRs 'refactoring CUDA' is planned.
- GSoC:
  - support for Xfeat: PR is raised
  - Disk+lightglue: in progress, PR will be submitted soon.
- Agrim:
  - stripped down G2O library is put into opencv/3rdparty: done locally, will be submitted soon. depends on visual odometry (see below).
  - visual odometry PR #29331: under review
  - loop closure and detection and global BA implemented.
- Akansha (BigVision intern):
  - PR #29332 optimized, cleaned up IPP code (moved to HAL)
  - PR #29374 - extracted IPP to HAL for resize (4.x).
  - PR in progress: Integrating OpenVINO as backend for the new engine
  - PR in progress: optimizing fastgemm and migrating it from dnn to core; need to support batch gemm and activations
- Kriti (BigVision-intern):
  - fixing grid sample (bicubic)
  - fixed opset-20 mode parsing.
- Need to continue working on documentation for a bit longer; and maybe put it back to the previous hosting where it can accessed more easily.
- Ayush:
  - migrating IPP to HAL
- Savya:
  - PR #29372: initial step for HIP integration
  - PR #29386: Working on dynamic quantization
- Varun:
  - working on extending CI
    - fixed bug in Mlas integration
    - new Windows CI builder is being prepared
    - adding priority to PRs to accelerate testing of some most important PRs
    - looking on how to bring testing time in windows down.
- Om Rope:
  - working on COOL (OpenCV for AWS).
    - average acceleration over regular OpenCV 5.0 is 1.2x. 100% of accuracy tests pass.
    - the next step is to test opencv-python 5.0.
Shiqi Yu (OpenCV China):
- several students now contribute to OpenCV: 10 PRs are still pending.
- some students work hard on RVV optimization.
- all students use AI in different degree, that makes it difficult to evaluate students, but that's inevitable.
Vincent:
- slowly switching to OpenCV 5.0, module-by-module.
Phil (opencv.org):
- proposal for AWS-funded competition is being prepared.
- talking to Qualcomm on the next possible contract.
  - Qualcomm-organized competition is also in plans.
- collaboration with AMD is going on, quite successfully.
- new OpenCV T-shirt is released! https://opencv.myspreadshop.com/?idea=6a39b4d464a4496c98165465
- there was quite interesting conversation with NVidia at CVPR.
- thinking of releasing a series of small books on OpenCV 5.
Alexander S (opencv.ai):
- working on several tracks:
  - optimization track: there are multiple PRs from community
    - incoming stream of PRs from AMD folks with AVX2/AVX512 optimizations on operations with mask, cvt color etc.
    - PRs from SpaceMIT with RVV optimization. E.g. resnet50 now runs 10x faster than in 5.0 (still not quite fast on K1, hopefully it will be faster with K3)
    - Alexander tries to verify all the claimed optimization results on our hardware (x64, ARM, RISC-V)
  - non-CPU HAL: several PRs on adding CUDA, HIP, Metal-based extensions.
  - another track is with many small patches to 'fix' complains from various sanitizers. Sometimes they are useful, sometimes they are not.
  - CI track:
    - trying to add CI configuration with musl-based linux
    - build/test acceleration. Sometimes builds are not cancelled timely, so we have unnecessary duplicated jobs when some PRs have been updated but their older commits are still in the queue.
  - working on opencv-python package release. There is a request to rename it to 'opencv' (so that users just run pip install opencv to install opencv-python).

2026-06-03

Vadim: *
Abhishek Gola & the team (BigVision):
- Documentation:
  - we moved to AWS for now for hosting docs.
  - brand-new OpenCV 5 docs with the modern look-n-feel will be put there tomorrow, links will be updated.
- Working on extra assets for OpenCV slides for CVPR.
- VLM sample will be merged today, wiki will be updated tomorrow.
- DNN performance table will be put into wiki tomorrow as well.
- There is some update on SLAM: bundle adjustment has been integrated, loop closure will follow.
Vincent:
- managed to compile OpenCV 5 internally.
- waiting for 4.14 as well.
- need to get rid of alloca.
Shiqi Yu (OpenCV China):
- students submitted a few PRs, a few of them have been merged.
- RVV optimization work is one of priorities, may get some extra (not huge) funding to do that.
Alexander S (opencv.ai):
- Final preparations for OpenCV 5.0 release.
- DNN:
  - most of the patches have been merged.
  - trying to finalize PRs with DNN features
  - Caffe cleanup is finished, PR is merged
- VideoIO module:
  - PR with return values is merged
  - ffmpeg wrapper PR is in progress
- Final 3D functionality reorganization:
  - we are going to have not just geometry module, but also ptcloud.
    - geometry will depend just on core
    - ptcloud is going to be much higher-level module, depending on geometry, imgproc, potentially dnn too.
- Steps towards the release: C++ package by friday, python package will be released next week.

2026-05-27

Vadim:
- working on more accurate and faster implementation of image warping functions with bicubic interpolation
- also working on faster int8 convolution
Abhishek Gola & the team (BigVision):
- Abhishek:
  - raised PRs:
    - #29161 (fixed conv+add+batchnorm)
    - #29104 (SPDA fusion layer)
    - #29091 (Sphinx-based documentation)
  - merged PRs:
    - #29076 (moved GPU core changes to separate PR)
    - #29098 (block-layout channelsPReLU)
    - #29099 (GpuMat CI warnings)
    - #29126 (attention graph fusion)
    - #29087 (fixed write status in videoio)
    - #29073 (DISK feature extractor)
  - GSoC:
    - aliked + lightglue PR #28986
      - added dnn-based panorama stitching
    - merged 2 PRs
    - next steps: clean aliked + lightglue, submit DISK+lightglue
- Kriti (BigVision intern):
  - Completed updating tutorials, python tutorials, javascript tutorials, introduction, faq, bibliography
  - verified documentation formatting
  - getting prepared for reference part update.
- Agrim (BigVision intern):
  - visual odometry is in codebase.
  - added new in-tree opencv slam module: monocular visual odometry, two-view bootstrap, PnP, keyframing + map growth, COLMAP trajectory, kyframes etc.
  - loop closure & bundle adjustment is not there yet.
- Ayush (BigVision intern):
  - PR #28912 (darknet cleanup)
  - PR #29127 (KV-cache)
  - PR #1355 (darknet cleanup in opencv-contrib)
  - PR #29012 documentation improvements.
  - PaliGemma goes out of memory bacause of duplicate memory buffer allocation by MLAS.
- Naresh (BigVision intern):
  - Benchmarked several more models (DINO, SAMURAI, TrOCR). Now we compare even better to ORT. 26% degradation in the worst case (DINO), SAMURAI is 46% faster!
- Varun:
  - merged:
    - fixed int32 overflow (when computing shapes) in DNN #29093
    - #28945 perf test 14 models
    - added missing data layout constant in Java bindings
    - int8 optimization
    - relax ViT_B_32 llinf threshold
  - raised:
    - reexported YuNet with proper dynamic-shape inputs
  - working on:
    - adding --download-large-model flag to download_models.py script.
- Om Rope:
  - AWS-COOL:
    - working on COOL Graviton 4 ver 3.
    - optimized some key functions:
      - DCT/DFT/HoughLinesP: 2x-3x acceleration
  - Working on documentation (tutorils): PR #29100
Vincent:
- trying 4.14-pre locally
- truco contours apparently gives a different result sometimes.
Gary:
- no big updates
- talking to Edgar Riba
Phil (opencv.org):
- CVPR is very close
- going to meet with Intel at CVPR
- preparing assets for OpenCV 5 launch (maybe including some videos with OpenCV users' success stories)
- preparing preview of OpenCV 5 features
- communication with Qualcomm goes on
Shiqi Yu (OpenCV China):
- students are working on various OpenCV stuff (RISC-V, dnn feature detectors, etc.)
Alexander S (opencv.ai):
- Plan for OpenCV 5:
  - do feature freeze this weekend
  - next week stabilize the package
- Looks like some regressions have been introduced in 4.x recently.
- There is a strange failure of camera calibration that is reproduced on one server in OpenCV China, but locally it cannot be reproduced.
  - Maybe the problem relates to some AVX512 code paths being executed.
- Merged most of DNN patches
- Merged InputArray(vector<vector<T>>) patch from Vadim
- Prepared patch to migrate lot's of 3d and 2d geometry stuff to dedicated module geometry. '3d' module will disappear.
- Caffe removal is in progress (will be finished before release)
- Another polynomial equation solver has been submitted. It works quite well!
- Many RISC-V patches have been submitted.
- Circle Detector has been accelerated substantially.
- Maksim Shabunin has tweaked our PR submission hooks so that AI-submitted patch have a special 'robot' mark.

2026-05-20

Abhishek Gola & the team (BigVision):
- Abhishek:
  - PR #29076 (moved GPU core changes into separate PR)
  - PR #29073 (added DISK feature)
    - issue: #27083
  - merged PR #28971 (subgraph name scoping)
  - merged PR #28752 (net profiling)
  - working on another PR to accelerate attention (speed is quite good)
  - Feature extractor and feature matcher:
    - aliked + lightglue PR #28986
    - going to submit: added optional MaxPool output
    - going to submit: fixed batch size issue in Gather
    - going to start working on: add dnn-based panorama stitching
- Kriti (BigVision intern):
  - Working on documentation for OpenCV 5.x
  - Modules updated:
    - Core
    - DNN
    - Imgproc
  - Verified documentation builds, removed old/stale parts
- Agrim (BigVision intern):
  - working on modular monocular SLAM for OpenCV
  - started with the odometry part
  - loop closure part will be added later
- Ayush:
  - working on KV-cache: modifieid KV-cache implementation, PR is going to be submitted soon.
    - we are much closer to ORT-genai speed for now.
- Naresh:
  - benchmarked a few more models: ORT processes SwinIR ~3x faster than OpenCV, but for other models the difference is smaller or OpenCV is even faster.
- Varun:
  - PR #29074. added datalayout constants to Python
  - PR #29079: YuNet int8 optimization: 4.39=>4.01ms optimization, 4.52=>4.32ms (per channel)
    - dedicated 1x1 convolution
    - QLinearMul support
    - Broadcast shape support
    - block layout propagation
  - Alpine Linux (with Musl-C) has been added
  - Researching difference in performance between still image and video processing speed with DNN
- Om Rope:
  - Working on AWS-COOL, Graviton 4 version 3 support.
    - optimized 3 functions:
      - ORB: accelerated y 14.4%
      - StereoBM: 10%
      - FastNIMeans: 10.6%
  - PR #29078: optimizing FP32<=>FP16 conversion with CUDA
  - PR #4122 (opencv_contrib): working on updating cudev tests.
Vincent:
- Tried OpenCV 5
  - see a few accuracy-related tests, but nothing exceptional
- submitted PR to unify calib3d API in 4.x and 5.x.
Shiqi Yu (OpenCV China):
- teaching C++ course. Students are encouraged to submit some PRs to OpenCV, 10 PRs have been merged.
  - possibly, some of those students may become OpenCV interns (?)
Phil (opencv.org):
- preparing for CVPR
- going to have meetup with Intel at CVPR
- contacted with UNIKE (?) from Netherland regarding possible collaboration
- also contacted with '3D display' company, who may also join opencv.org as bronze members
- 5.x demo is being prepared for CVPR
- going to have live stream 'OpenCV 5 preview' at May 28
- talked to some cool project (https://www.kickstarter.com/projects/backyardbrains/spikerbot-build-a-brain-create-a-creature) from Michigan Univ.
- waiting for some financial numbers from OSCCA organizers — going to take part in OSCAA'27.
Gary:
- talking to Edgar Riba on spatial AI stuff.
- work with/at Stanford is going to start in June.
  - probably will work there on 'JEPA' models by Yann Lecun.
Alexander S (opencv.ai):
- Working on incoming PRs
- Working on the final (for 5.0) reshuffling of functions:
  - 3d + part of imgproc => new module geometry
- Working on various camera calibration improvements
- Active work on DNN continues (thanks to BigVision team):
  - optimization
  - cleanup (Caffe and Darknet support will be removed)
- There is a series of PRs about RISC-V from Shiqi's teams and not only.
- Nearest plans:
  - finalize another 4.x to 5.x merge
  - after that merge as many PRs on DNN as possible
- Slight Input/OutputArray refactoring in core part is still pending (PR should be ready soon)
- CI updates:
  - had problems in Linux builders (constant disk overflow)
    - found the root cause and fixed it.
  - Alpine linux builder now passes all tests.
    - going to add this configuration to CI
  - Ubuntu 26.04 configuration is not ready yet, there are some test failures and also non-trivial warnings.
  - Windows-on-ARM is there, but not enabled by default, there are also some test failures.

2026-05-13

Abhishek Gola & the team (BigVision):
- Abhishek:
  - raised a few PRs, including #28971 (subgraph name scoping)
  - merged a few PRs, #28966, #28964 (extended primitive core functions to support new data types), #28963 (custom layer support), #28957 (extended attention), #27560 (conv transpose)
  - finalized 2 more PRs: #28934 (MLAS), #28966 (dequantizelinear slice bug fix)
  - GSoC: finally have PR #28986 (aliked + lightglue). There are still a few TODO items:
    - add support for the optional second output to MaxPool
    - extend Power operation
    - add dnn-based panorama stitching
  - Kriti (bigvision intern) is working on revising OpenCV docs feel & look. It's using a combination of doxygen and sphinx. The current version is incomplete, but looks promising
  - there is an option to put more developers to work on CI. What they could work on?
    - answer from Alexander S: Windows is definitely priority no 1, since currently Windows configuration in CI is unstable; we also don't support well Windows-on-ARM. RISC-V is also important (less important than Windows)
- Ayush:
  - merged PRs #28968 (0D in dnn), #28982 (unconnected out). With MLAS performance of RF-DETR, RT-DETR and Grounding-dino_tiny matches or exceeds ORT.
- Naresh:
  - Measured RTMPose-m, U2-Net, Real-ESRGANx4. ORT is 10-20% faster than the new engine.
  - SwinIR and BiRefNet fail because of buggy Slice. Should be fixed this week.
- Varun:
  - merged PR #28969: updated ORT to 1.25.1
  - merged PR #1359 to opencv_extra (18 ONNX models)
    - need some mirror for HF accessible in China, e.g. hf-mirror.com or https://www.modelscope.ai/home
  - optimized YuNet int8 model. 7.32ms=>4.52ms
    - fused sigmoid, maxpool, residual add
    - the new int8 model is also more accurate.
  - working on adding Ubuntu 26.04 to CI runner.
Jia Wu (OpenCV China):
- currently warpAffine/warpPerspective in OpenCV 5.0-pre on RISC-V is obsolete, need to update accelerated kernels. The students from Shiqi's group are looking on how to fix it.
- the new DNN engine does not work properly on RISC-V + RVV, many tests fail.
Phil (opencv.org):
- OSCCA conference is over, OpenCV events attracted many people. So, it was very successful.
  - also, had a booth there.
  - with such a success, planning to repeat it next year.
- EVS is also over.
  - again, had a booth there.
  - Satya is giving 2 workshops today.
- CVPR is really close.
  - will have a booth there.
  - working to produce some give-out stuff for it (mugs, stickers etc.)
- Talked and continue talking to several companies to become OpenCV sponsors.
Alexander S (opencv.ai):
- Done:
  - Ubuntu 26.04 is added to CI
    - a few test failures, related to gstreamer and camera calibration
  - working on adding Alpine Linux to CI with musl runtime library
  - very significant progress with DNN: merged many patches from Abhishek and the team
  - looking on the final TODO list for OpenCV 5.0
    - moved intelligent scissors functionality from imgproc to photo module.
    - working on finalizing geometry module that will be made of the current 3d module, parts of imgproc etc.

2026-04-29

Abhishek Gola & the team (BigVision):
- Abhishek:
  - PR #28811, #28889, #28859 merged (QLinear, SIMD kernel optimization, attention fusion)
  - working on:
    - optimized 3x3 depthwise convolution and fusion of conv+pad.
    - re-quantization of yunet-int8.onnx (the existing model is buggy)
    - working on Winograd-based convolution, have some promising results
  - GSoC:
    - aliked+lightglue API will be built on top of Feature2d + DescriptorMatcher
    - ran xfeat sucessfully as well, but it's not combined with lightglue yet.
- Ayush:
  - working on PR #28886 (ort-genai wrapper)
  - working on removing Darknet importer, replacing all Yolo models with onnx versions
- Naresh:
  - benchmarks have been extended to cover BERT, Vit, SegFormer B0, SigLIP Vision Encoder, Grounding DINO. We are getting closer to ORT, but in other cases we are still 2-3x slower than ORT.
- Varun Jaiswal:
  - Gemma 3 tokenizer has been merged
  - PR #28872: OpenCV DNN shape inference fix when ORT backend is used.
  - Looking on how to add Ubuntu 26.04 to our CI.
- Om Rope:
  - Merged PriorBoxLayer PR #28855
  - Added LpNormalization PR #28888
  - Updated WeChat QR model (PR #29746)
  - Updated colorization model (PR #28722)
  - Updated Python script in opencv_extra for WeChat.
  - Updated colorization model in opencv_extra
Oleksandr N (opencv intern):
- still working on kv-cache Paligemma. Get incorrect results, trying to debug it.
  - going to finalize the PR this week.
  - starting new job next week, will inform whether can continue work on OpenCV.
Phil (opencv.org):
- working on the approaching OpenCV conferences. The tickets are still available.
- Talking to Qualcomm and Amazon about the next OpenCV competition.
Shiqi Yu (OpenCV China):
- the new course has just started, students will study C++ and may probably contribute to OpenCV.
  - PR with ffmpeg bindings bugfix has been prepared by one of the students and was merged.
  - there are DNN+RISC-V+RVV issues. Maybe the students from Sustech can look at it:
    - #28870: https://github.com/opencv/opencv/issues/28870
    - #28852: https://github.com/opencv/opencv/issues/28852. The issue was already reproduced.
- Spacemit K3 may become available in May. We could possibly use it for RISC-V+RVV CI build configuration.
Gary:
- will likely start some research projects at Stanford.
Alexander S (opencv.ai):
- Worked on reviewing multiple PRs, mainly about DNN.
  - OpenCV China CI cannot fetch models from HF, need to download those models manually.
  - With ORT backend for some reason some tests fail, we need to debug and fix those issues one by one, because reasons are different.
  - ORT-genai PR is still in progress, the architecture needs to be finalized.
  - Ubuntu 26.04 has been released. Going to do extensive testing with it.
  - Still working on the module reorganization: 3d->geometry etc.
  - Tried OpenCV on Spacemit laptop; mostly it runs fine, but the new DNN engine needs to be fixed for RVV.

2026-04-22

Abhishek Gola & the team (BigVision):
- Abhishek:
  - submitted #28811 (QLinear), #28859 (attention fusion)
  - merged #28821, #28807, #28808, #28750, #28741
  - next step: double transpose->transpose, split->n conv->concat.
  - now we are faster than ORT on a few YOLO variants (on Intel platform with AVX2-VNNI)
  - GSoC: Feature2D and DescriptorMatcher will be the base classes instead of custom base classes.
    - Successfully ran Superpoint, Xfeat and Aliked feature extractors using OpenCV DNN. Decided not to include Superpoint support because of potential legal issues.
  - Continue to work on COOL (cloud-optimized OpenCV), version 2 has been released
- Varun Jaiswal:
  - Bug PR #28816: fixed Flatten node
  - Gemma3 tokenizer support in dnn (PR #28837)
  - Bug PR #28836, GELU SIMD
  - Softmax: -inf inputs produce sum=0.
  - testInputShapes(). miscount number of inputs.
- Om Rope:
  - AWS-COOL v2 is now publicly available.
    - 3.22x speedup across top 2 functions.
    - all optimized functions pass accuracy tests.
  - caffe importer cleanup.
    - PR #28855: working on PriorBox layer to support WeChat QR model
- Naresh:
  - mobilenet, yunet are still slower (2-3x slower), but some models are already faster than ORT.
  - Alexander S, Abhishek: it seems that yunet 8-bit does not provide good accuracy
- Ayush:
  - PR #28799 merged (more models added to dnn perf tests)
  - pad2, reshape2, resize2, shape, instancenorm and groupnorm extended to supported block layout
  - tokenizers: working on adding LLM class for the tokenizers
Oleksandr N (opencv intern):
- submitted patch with KV-cache. Going to test it with Paligemma, together with tokenizer that now supports Gemma 3.
  - for now the postprocessing is very simple but can be made more complex in future.
Vincent:
- continue to integrate OpenCV 5. only 300 tests are failing in the internal test suite.
  - there is issue with calib3d being replaced by several modules, which is a problem for Java users.
  - also, some functions behave completely differently, like calibration or mcc detector.
Shiqi Yu (opencv china):
- no big updates
- the group of students who work on RISC-V+RVV will probably take a look at DNN test failures.
Phil (opencv.org):
- working on the conferences (which coming really soon)
Gary:
- xfeat and aliked are really free alternatives to superpoint, so let's use them instead.
Alexander S (opencv.ai):
- Mostly focused on reviewing and merging PRs, especially DNN patches.
- Submitted issue about RISC-V+RVV-related issues in DNN: #28852
- Found problem with combination of IPP + CLapack + AVX2 baseline. Working on the fix
- Made some more fixes in multi-view calibration
- Continue refactoring/reshuffling of some calib+3d functionality:
  - calibration rig detectors are moved to objdetect.
  - 3d will be renamed to geometry and will get some functionality from imgproc (convhull, subdiv2d etc.)
- Made more experiments with KleidiCV
  - looks like too many tests in DNN fail because of differences in resize.
- Enabled new pipeline in CI to build and run OpenCV without IPP.

2026-04-15

Abhishek Gola & the team (BigVision):
- Varun Jaiswal:
  - bug in Flatten: fix submitted PR #28812, opencv_extra PR #1342.
- Om Rope:
  - importer (Caffe) cleanup:
    - majority of the models have been converted to .onnx
    - pose detection, ssd vgg object detector and a few other models are still to be converted or replaced with onnx alternatives.
  - working on Graviton 4 COOL
- Ayush:
  - PR #28799 added more models to perf tests
  - working on adding block layout support in resize, pad, instance normalization, shape.
- Naresh:
  - MobileNet v2 1.5ms, 107.42ms ViT, MobileVit_XS 21.93ms, MobileVitv2_100: 7.93ms
  - going to measure SAM and Yolo's.
Oleksandr N (opencv intern):
- resumed work on KV-cache. The first goal is to enable KV-cache in FP32 precision. PR will be submitted soon.
- BigVision team will work on Paligemma support in our tokenizers.
Vincent:
- tried OpenCV 5.
  - have some build issues, but they are minor.
  - have some accuracy issues because of different semantics in some functions.
  - there is issue with int64 compatibility (some universal intrinsics are missing) in emscripten (wasm).
- applied for open-source grant internally, but it did not work. There is still hope, probably, need to apply slightly differently.
- C API LevMarq is working better than C++: github.com/opencv/opencv@934c0716d2/modules/calib3d/src/calibration_base.cpp (L1397)
Shiqi Yu (opencv china):
- 1st group of students: working on lightglue and features. Abhishek introduced progress last time.
- 2nd group working on RVV optimization: laplace is being optimized. PR will be submitted in 1 or 2 weeks.
- 3rd group working on QR detectors: there is some good progress, but the API and code structure needs to be reviewed.
- on SLAM: again, the progress is good. It's going to be a PR to opencv_contrib (new module?) PR #4101.
Phil (opencv.org):
- OSCCA (https://opencv.org/oscca) is very close:
  - linkedin post: https://www.linkedin.com/feed/update/urn:li:activity:7447374742518382592
  - x post: https://x.com/opencvlive/status/2041608025849442696
  - selling tickets. Students have some special discounts and there is certain number of free slots (20 slots) for students: https://form.jotform.com/260965697047068
  - sponsors are all there.
- partnering with Tangram: https://opencv.org/tangram-vision-and-opencv-are-partnering-to-fix-your-calibration-problems/
- Youtube channel gets more and more views
Gursimar:
- meeting AMD next week
- continue conversation with Qualcomm.
  - MS provided us (all open-source projects) with free virtual machines with Windows @ ARM. Unfortunately, the toolchain there (compilers) is not super-stable/mature, we experience issues here and there.
  - also have laptop with Qualcomm chip, but it also does not work well.
- documentation: someone can start looking at it in May.
Gary:
- will likely get invitation to Stanford robotics and hope to do some cool stuff there.
- dealing with some new startups
Alexander S (opencv.ai):
- merged many DNN-related PRs from Abhishek and team. DNN progress is quite good.
- working on reshuffling some functionality between modules before OpenCV 5.
  - so far, we agreed to move calibration rig detection to objdetect
  - we will rename 3d to geometry and move there some computational geometry stuff from imgproc
  - calib module will still be there, as well as stereo.
  - some functionality, like point cloud I/O, may require a dedicated module.
- working on Python bindings to support the latest Python versions, the latest numpy versions etc.
  - in particular, lock-free version of OpenCV Python package is requested by some users
- going to enable ONNX conformance tests to measure our ONNX coverage on continuous basis

2026-04-08

Vadim: *
Abhishek Gola & the team (BigVision):
- Abhishek:
  - Activation fusion #28750: in progress
  - Net profiling #28752: in progress
  - ORT w. GPU: #28588 merged
  - int8 block layout #28741: in progress
  - loop layer #28121: in progress; the test has been added
  - adding eyelike layer: PR is being prepared
- GSoC:
  - SuperPoint.onnx support: draft PR submitted, it's being polished.
    - there is repo that claims to reimplement SuperPoint under MIT license: https://github.com/rpautrat/SuperPoint?tab=MIT-1-ov-file. Need to look more closely at it.
  - #28738: sub-modules for 'features' module to do feature detection and matching. Need to discuss whether it's a good solution to create submodules (nested namespaces) and whether deep learning features can fit into the same framework as we already have (to be compatible with image stitching etc.)
    - once it's done, Xfeat and ALIKED features will be integrated.
- Ayush Kumar:
  - bug fix in NMS #28749
  - #21031: cannot be reproduced anymore
  - loop layer #19977. it's still valid with the latest version of Loop patch. need to look further.
  - telea inpating issue
- Naresh Raja:
  - added Depth Anything v2 and SAM2 into DNN benchmark
    - we are 1.6 to 2.7x slower than ORT
    - with the latest patches pending hopefully the performance will improve even more.
  - (note from Alexander S and Abhishek): going to add some of the models into performance tests.
- Varun Jaiswal:
  - added Qwen 2.5 tokenizer support
    - with fix for Flatten
- Om Rope
  - batchnorm before conv (or after relu): still pending
Vincent:
- working on resolving some fuzzing issues.
- OpenCV was internally tested on ARM running Windows: good news
  - message from Alexander S: we are close to adding Windows ARM as CI configuration, but it's still not very smooth — some tests are failing. But it compiles in default configuration.
- request: please, inform me about OpenCV 5 more-or-less-ready-to-test version.
Shiqi Yu (OpenCV China):
- several groups of students are working on different topics, but no big updates yet.
Phil (opencv.org):
- Quite tough (yet exciting) schedule in the coming 2 months:
  - OSCCA is approaching: https://opencv.org/oscca/, https://www.linkedin.com/feed/update/urn:li:activity:7447374742518382592.
  - EVS will follow OSCCA
  - CVPR will follow EVS
- Talking to Tangram Vision, we are now partners: https://opencv.org/tangram-vision-and-opencv-are-partnering-to-fix-your-calibration-problems/
- Continue conversation with Qualcomm, but it's going slowly.
Gary:
- Contacting to some old friends at Stanford regarding possible joint projects.
- Thinking of some auxiliary "OpenCV Pro" version, same base (opencv repo), but extra services (+ maybe a few add-ons to solve some niche problems).
  - e.g. multi-room scan. There are some interesting approaches that seem to do it well.
Alexaner S (opencv.ai):
- working on resolving multi-camera calibration failures after recent patches
- working on preparing DNN for OpenCV 5 release
- ONNX conformance test suite. We are trying to adopt standard ONNX scripts. Apparently, ONNX team is already measuring OpenCV compliance ratio, but they test the latest official OpenCV release. We just need to add support of OpenCV5-pre to this already created framework for OpenCV + ONNX.
- DNN tokenizers. Qwen 2.5 is now supported.
- Need to look at the module layout once again. Maybe we should revise our decision to split calib3d into parts before 5.0.
- Looking at the latest Aruco patches (actually, submitted during the past GSoC). Going to polish and finally integrate them. Also want to extend AruCo test suite significantly.

2026-04-01

Vadim: *
Abhishek Gola & the team (BigVision):
- PR #28741 submitted (int8 conv for block layout)
- initial qdq support: patch merged
- conv 1x1 + 3x3 acceleration: patch has been merged
- convtranspose PR: working on a completely new implementation compatible with block layout
- tested different parallel frameworks on resnet50-qdq.onnx
  - pthreads is the fastest (faster than OpenMP or TBB)
- feature extractor and matcher: debugging superpoint.onnx inference with OpenCV DNN
  - initial superpoint and lightglue support PR is being prepared
- fixed Loop layer (PR is pending) to allow inputs from outside of subgraph
- NMS-Reshape layer: issue #20258 is solved, PR will be submitted
- ORT-GenAI GPU backend: PR will be submitted once ORT-GenAI patch is merged
- Done initial version of OpenCV 4.x DNN vs 5.x DNN vs ORT benchmark. Usually OpenCV 5.x is noticeably faster than 4.x, but still slower than ORT, sometimes much (e.g. 4x) slower.
- added Qwen 2.5 support to OpenCV tokenizer. Adding more tests now.
- resolving ORT-related minor CI issues (compiler warnings).
- replaced wechat qrcode with onnx model:
  - added priorbox parser to support wechat model
  - relaxed nary-elemwise shape inference restrictions
- added necessary onnx models to opencv_extra.
Phil (opencv.org):
- 1 month away from OpenCV conf (OSCCA) at LA, after that there will be EVS (Santa Clara) (11-13th of May).
  - Satya will give talk and trainings about VLMs.
- preparing for CVPR 2026
- talking to some AI chips manufacturers.
- getting new OpenCV sponsors.
- talked to BigVision management, who are now big supporters of OpenCV (big part of dev team)
- there has been interesting conversation with producers of F1 cars.
- talking to a producer of SoTA camera calibration solutions.
- going to have some active communications with Orbbec after conferences in May-June.
- also going to have conversations with Qualcomm during the Summer.
- OpenCV is now at Patreon
- OpenCV for AWS (a.k.a. 'Cloud-Optimized OpenCV Library' or COOL) is out, hopefully we'll get some money from it.
Gary:
- some conversations with more Shenzhen-based camera producers etc.
Oleksandr N (OpenCV intern):
- finalizing and testing GSoC tokenizer. Fixed bugs, now it's working properly.
- working on Paligemma tokenizer.
Alexander S (opencv.ai):
- ARM has released the latest version of KleidiCV. Tested it with OpenCV and it works well.
  - there is minor issue in one of DNN tests
- working with Qualcomm, who are working on Windows @ ARM. OpenCV now more or less compatible with this configuration
- merged several patches from Abhishek and the team regarding DNN
- preparing for another 4.x=>5.x merge.
- also working on improving CI infrastructure, e.g. adding the address sanitizer.

2026-03-25

Vadim: *
Abhishek Gola (BigVision):
- conv 1x1 and 3x3 have been accelerated (#28691)
- engine and backends params added to dnn examples (#28708)
- int8 support is about to be added (will be submitted soon). great acceleration on resnet50-int8 (40+ms => 6.6ms)
- block layout for deconvolution (#27560)
- API for deep features and matchers is in pgoress. superpoint model for now cannot be inferenced correctly, so using ORT backend for now
- there is good progress on integrating ORT-GenAI (testing on Paligemma model, both CPU and GPU (NVidia) paths).
  - found issue with Resize shape inference when using classic engine and OpenVINO backends.
- BlazeFace detector is 4.6x slower than ORT.
- CI pipelines improvements.
  - Added configuration OpenCV w/o IPP.
  - Found and fixed one issue in triangleRasterizeInternal.
  - CI check added to confirm that satellite PRs are submitted to opencv_contrib/extra (if needed).
- Working on Caffe importer removal in OpenCV 5 (PR #28678, contrib: #4094, extra: #1324)
- PR submitted to fix one problem in graph fusion: #28722
Shiqi Yu (OpenCV China):
- the 1st group of students are working on bringning LightGlue, deep features to OpenCV
- the 2nd group are working on RVV optimization: they will start with Laplace function.
- prof. from Spain is going to submit to OpenCV some accelerated algorithms (contour extraction?)
- the 3rd group are working on a better QR code detector
- another group are working on SLAM. Probably need some help to decompose the task and submit a series of PRs instead of one big PR.
Vincent:
- potentially found another source of funds for OpenCV
Gary:
- CVPR keeps growing; we will have a booth at CVPR 2026.
Oleksandr N (OpenCV intern):
- continue working on tokenizer, found some bugs in DNN inference of GPT-2 model
Alexander S (opencv.ai):
- working on PR review, in particular, on GPT-2 and tokenizers
  - found some problems in the current attention implementation
- more improvements on calibration
- planning to finalize and merge a few PRs to DNN.
  - after than we plan to submit patch to ONNX compliance committee to measure OpenCV ONNX compliance ratio.
- found that opencv_contrib's CUDA resize is faster than modern NPP resize.

2026-03-11

Vadim:
- Finally submitted patch with block layout-based convolution
Abhishek Gola (BigVision):
- PR #28635 (getFlops supported), #28637 (fix for the tracker models) submitted
- resolved comments on OnnxRuntime-GPU, as well as PR #28634, #28637
- next steps: qdq, block conv 3x3, 1x1, ...
- 3 GSoC-like projects started:
  - ort-ghenai
    - PR is submitted
  - comprehensive validation and sample generation for modern dnns
    - added sample for Yolo 26, currently (before block layout) runs >2x slower than ORT (45ms vs 20ms).
    - RT-DETR, RF-DETR
  - deep-learning based feature detector and matcher
- also working on improving CI:
  - fixing some bugs with CPU_BASELINE=AVX2
  - going to add Windows-on-ARM configuration (Qualcomm CPU)
- AWS-COOL (cloud-optimized OpenCV): achieved up to 25% acceleration on image analysis benchmark and up to 37% on image augmentation benchmark
- working on replacing .caffe models with .onnx models to get rid of obsolete Caffe format in OpenCV 5.
Oleksandr N (OpenCV intern):
- working on tokenizer. the existing GSoC patch supports GPT2 and GPT4. Adding support for PaliGemma would not be difficult to get.
- ort-genai based tokenizer: trying to build it locally.
Phil (opencv.org):
- still preparing for the conferences (EVS 2026, OSCCA 2026, GDC 2026, CVPR 2026, ...)
  - going to have more participants in OSCCA this year vs last year.
- in several weeks going to announce some new challenge
- Q from Alexander S: still not in the security mailing list.
Gary Bradski (opencv.org)
- no major updates, except for OpenCV summer of code. It's up and running (3 projects so far, see Abhishek report above).
Shiqi Yu (OpenCV China):
- 2 students working on LightGlue matcher.
- 2 more students will likely work on RISC-V optimization.
Alexander Smorkalov (opencv.ai):
- looking at several DNN-related PRs:
  - patch by Oleksandr N looks good, ready to merge it
  - looking at patch by Vadim (block layout), going to merge it soon as well
  - ORT-genai patch is under review as well
- working on further improvements of HAL

2026-03-04

Alexander Smorkalov:
- Debugged and fixed mul overflow for U16 #28579
- Manually tested ONNX Runtime integration for CPU #28444 (merged), works on GPU #28588
- Reviewed, tested, helped to resolve issues: ## 28592, 28591, 28590, 28587, 28584, 28578, 28576, 28575, 28574, 28570, 28569, 28548, 28289, 28309, 28536, 28519, 28582, 28589, 28585, 28508, 28482, 28355
Ivan Avdeev:
- Fixed Linux server outage in Sustech
- Helped BigVision team with new server deployment
- Works on Masl-C based test environment
Gary Bradski
- Will be funding OpenCV Summer of Code
- Since we have to focus on getting out release 5.0, we are going to focus on this for 2026

2026-02-25

Abhishek (BigVision):
- DNN: finished GRU layer
- ORT follow-up patch has been prepared, will be submitted after the first ORT patch is merged.
- Improved Reduce layer has been merged.
- 4.x=>5.x caused incorrect shape inference bug: submitted patch
- trying to setup more CI configurations (Linux-based):
  - setup another Linux-on-Xeon server that could run some CI configs
  - prepare several new Linux containers: ORT support, ...
- Abhishek's student may take a look at ORT-genai as the library to handle LLMs/VLMs.
Om (BigVision):
- 1st version of Graviton 4-optimized OpenCV has been prepared and is being tested. so, we observe:
  - On Graviton 2 = 17.6% gain.
  - On Graviton 3 = 20% gain.
  - On Graviton 4 = 23.8% gain.
Gursimar:
- no updates
Phil:
- still working on OpenCV conference stuff.
  - asking Gary to ask Ethan R to officially confirm participation.
- for EVS already have 4 speakers (including Ethan), also finalizing list of speakers for OSCCA.
- working with AWS (finalizing OpenCV on Graviton's) and also talking about partnership/sponsorship of OpenCV.org.
- GDC (Game Dev Conf) is approaching, will probably take part there as well.
Gary:
- since we missed GSoC this year, let's select 3-4 projects and run them by ourselves.
Oleksandr N (opencv intern):
- finalized patch with attention-w-kvcache-support kernels.
- will now look at the unfinished tokenizer #27534
Alexander S (opencv.ai):
- 4.x=>5.x merge is done, but not all the changes have been merged. Another round is expected next week.
- GTK 4 support has been contributed (#....), but the code seems to be AI-generated. Can be used as starting point.
- several patches have been contributed to improve DNN, mostly by Abhishek.
- met with Abhishek and Ivan to discuss the future CI improvements.
  - in particular, musl-based linux distro's should be added.
- continue discussions with ARM, Intel on the respective HALs (KleidiCV and IPP, respectively).

2026-02-11

Abhishek (BigVision):
- RoiAlign layer has been merged.
- looking at opencv issues related to DNN. 20 issues have been closed.
- PR on improved batch normalization has been submitted (#28522)
- ONNXRuntime patch (#28444) is being polished.
- looking on how to convert all Caffe models used in the tests to ONNX or TFLite.
Om (BigVision):
- Continue to optimize OpenCV for Graviton 4.
  - Working on CLAHE acceleration (1.56x speedup has been achieved).
  - 3x speedup (i.e. came from 3ms to 1ms) acceleration on adaptive threshold.
  - Looking at connected components acceleration.
Vincent:
- nothing to report
Gary:
- reviewed GSoC 2026 page once again.
- Kornia wants to join our GSoC program.
Phil (opencv.org):
- focusing on OSCCA 2026 conf; little reshuffling of the schedule.
- github secure funding program: contacting to github folks; in particular, this PR was mentioned: https://github.com/opencv/opencv-python/pull/1051
- considering to participate in GDC conference
Oleksandr N (opencv intern):
- working on the new kernels for paged attention.
  - submitted draft PR today (#28524).
Alexander S (opencv.ai):
- focusing on reducing tech dept.
- merged a few patches that improve/accelerate DNN.
- need to resubmit Vadim's patch with improved vector<vector> support to 5.x manually.
- found a few regressions (accuracy issues) with OpenCV + KleidiCV, exploring it now.
- planning to continue with 4.x => 5.x (#28523)
- fixed python bindings for Charuco boards #28517
- reviewed, tested, helped to resolve issues ## 28308, 28379, 28451, 28323, 28500, 28499, 28504, 28505, 28510, 28511, 28460, 28453, 28506, 28470, 28475, 28477, 28415, 28266, 28488, 28521
- CI: prepared CI development plan on Gihub. The main goal is to increase CI coverage, e.g.:
  - CI config with disabled HAL
  - DNN+ORT config (BigVision will work on it).
  - Alpine (Musl) linux with alternative C library.

2026-02-04

Gary:
- (could not take part in the meeting, but sent update): we applied for GSoC 2026, the results will be announced on Feb 18th.
Abhishek (BigVision):
- layernormalization has been merged.
- new graph engine is being designed: https://github.com/abhishek-gola/opencv/pull/21/changes. Waiting for the comments.
- looking at opened issues regarding DNN, trying to reproduce them. Many of the issues cannot be reproduced anymore with 5.x branch.
Om (BigVision):
- OpenCV for Graviton 4 activities. There are some regressions, working on them.
Oleksandr N (OpenCV intern):
- there is kv-cache PR (#28484), but it seems to be broken
- continue working on real kv-cache support, one of the steps (currently working on it) is to extend batch gemm to operate on dis-contiguous memory storage.
Vincent:
- sandboxing video i/o in OpenCV?
- playing with some opencv 5 changes locally.
Phil:
- working on OpenCV conferences (OSCCA, EVS, ...)
  - got confirmations from some speakers
- starting to promote cloud-optimized OpenCV; there is even special logo for it.
- Intrinsic has become OpenCV sponsor again.
- thinking of the next OpenCV competition, talking to potential partners.
Alexander S (opencv.ai):
- caught a cold, paused work for 2 days.
- couple of issues in arithmetics functions on arm
- need to discuss some more gsoc ideas
  - what about DICOM support?
- new opencv-python binaries for opencv 4.13 will be released soon.
- plan to focus on incoming PRs
  - in particular, a few interesting PRs have been contributed by one contributor (and these patches are likely LLM-generated):
    - YAML 2.0 support
    - initial DISK feature detector support. At least, we could import DISK model.

2026-01-28

Vadim:
- almost finished block-layout PR, just 5 tests fail.
Abhishek (BigVision):
- PR submitted to add RoiAlign layer
- PR: use OnnxRuntime as alternative engine for OpenCV DNN (#28444). Only Linux is supported.
Om (BigVision):
- working on Graviton 2 and Graviton 3-optimized builds of OpenCV.
- started looking at Graviton 4-optimized build of OpenCV.
Oleksandr N (OpenCV intern):
- working on KV-cache. Next week will provide perf numbers
Phil:
- working on OpenCV SID conf (2026 May); talking to potential speakers.
- working on OpenCV @ Embedded Vision (also 2026 May).
- OpenCV takes part in CVPR 2026.
- speaking to the existing and potential OpenCV sponsors (Intrinsic, OpenMV, ...).
- looking for some ways to get free advertising for OpenCV.org.
Vincent:
- no updates
Alexander S (opencv.ai):
- Done:
  - Alexander Smorkalov:
    - Plan to rebuild opencv_ffmpeg to address some users' complains.
    - Working on enabling AI-based automatic reviewing robot (that will just provide suggestions and will not perform any actions, of course).
    - Debugged and fixed extra OpenCV-Python-Headless dependency on X server #28438
    - Reviewed, tested, helped to resolve issues ## 28461, 28346, 28380, 28470, 28386, 28467, 28255, 28301, 28454, 28422, 27988, 28439, 28457, 28401, 28303, 28446, 28317, 28441, 28451, 28442, 28423, 28355, 28323
  - Ivan Avdeev:
    - Fixed infrastructure issues triggered by China infrastructure outage.
    - Works on MacOS Intel and iOS build failures (miss-configuration).

2026-01-21

Vadim:
- working on the block-layout PR, most tests pass, just ~30 tests fail.
Abhishek (BigVision):
- graph redesigned has started.
- exploring the use of ORT as alternative DNN engine (with OpenCV DNN-compatible API on top).
- need to setup a call regarding CI improvements. BigVision can try to replicate OpenCV CI, just with 1 server and 1 build configuration, in their private fork of OpenCV in a local github account.
- AWS activities: OpenCV for Graviton 2 went public yesterday. OpenCV for Graviton 3 is expected soon (likely, next week).
- Suggestion from Vadim: enable "-Ofast" instead of "-O3" in the core module; submit PR, see how many tests fail.
Gursimar (BigVision):
- CI updates:
  - the new person will be available starting from February who will work on CI
  - Intel workstation is available for the first experiments with CI.
Om Rope (BigVision):
- new engineer, working on AWS, specifically, on optimization of core and imgproc functionality.
Shiqi Yu (OpenCV China):
- no updates so far.
Oleksandr N (OpenCV intern):
- Whether attention PR can be merged? It would simplify the work on KV-cache.
Vincent:
- OpenCV 4.13 fully deployed internally; just a few tests (in opencv-based solutions, not in opencv itself) fail, so that needs to be addressed.
- Some patches for OpenCV to be compliant sandbox2 and other security tools will likely be submitted soon.
- From time to time OpenCV image codecs throw exception, but diagnostics needs to be improved.
- Q from Alexander regarding rotatedRect patch (#28381). A from Vincent: it really helps to improve precision when coordinates are big values.
Phil (OpenCV.org):
- just got back from OpenCV F2F forum.
- had recent conversations with Qualcomm, trying to setup a formal contract.
Alexander S (opencv.ai):
- Done:
  - Prepared release 4.13.0
  - Prepared materials for F2F in Shenzen
  - Works on AI submission guidelines
  - Reviewed, tested, helped to resolve issues ## 28380, 28439, 28431, 28426, 28425, 28423, 28419, 28410, 28406, 28404, 28401, 28422, 28308, 28340, 28355, 28375, 28381, 28441, 28292