mirror of https://github.com/opencv/opencv.git synced 2024-11-23 18:50:21 +08:00

Table of Contents

2024 OpenCV Technical Committee
Template

2024-mm-dd
Agenda

To Dos

Minutes

To Dos

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

2024 OpenCV Technical Committee

Meeting_notes

Template

2024-mm-dd

Agenda

To Dos

Name
- todo

Minutes

2024-11-20

Yuantao F (OpenCV China):
- Yuantao F:
  - submitted PR with performance regressions in warpPerspective, warpAffine & remap (merged).
  - working on nearest-neighbor interpolation flavor of warping functions
- Jia W:
  - reviewed samples by Gursimar. all the found issues have been merged.
  - parallel Annoy is made deterministic. Now need to evaluate quality.
Vadim:
- helped to fix LSTM
- removed G-API module from the main repo; to be moved to opencv_contrib.
- fixed Python detection on macOS (not merged yet)
- invited A. G. to present 'OpenCV @ HF' initiative.
Vincent:
- C++ calibration ported from 5.x back to 4.x.
- Observed degradation in calibration accuracy (C++ version of Levenberg-Marquardt solver in 4.x is worse than the old C implementation, but C++ implementation of LevMarq in 5.x should be better than both).
Shiqi:
- The new Wechat QR code: just finished evaluation. Will prepare PR within 1-2 weeks.
- Now preparing program for OpenCV F2F; the participants should provide titles and abstracts by Nov 30th.
Gursimar:
- finalized tracker and fixed bugs in a few samples.
- object detection. It uses some API that is not implemented in the new engine.
Phil:
- working on the new competition. Looking on how to employ container (docker) technology to make it safer for evaluation.
- "special AI conference" in 2025 is another event that opencv.org team is working on.
Alexander S (opencv.ai):
- Done:
  - Abduragim Shtanchaev:
    - Finished LSTM for new graph 🎉 #26391
    - Working on running diffusion model on new graph engine to check if dynamic shape are supported
  - Alexander Smorkalov:
    - Prepared 4.x->5.x merge #.
    - Dropped C API in Core module #26458
    - Tested and benchmarked FastCV integration as HAL #26316
    - Updated CI for G-API drop.
    - Reviewed, tested, helped to resolve issues with #26478, 26469, 26468, 26466, 26462, 26461, 26460, 26459, 26455, 26454, 26485, 26484, 26316, 26391
  - Alexander Lyulkov:
    - Working on modifying Tensorflow parser for the new dnn engine.
Vladimir D:
- removed the remaining traces of C API from 5.x
- updated ffmpeg wrappers for Windows.

2024-11-13

Vadim:
- [OpenCV China] fixed performance regression in warpaffine, warpperspective and remap when processing U8C4 data. PR has been merged. We now try to apply the same optimization for C4 cases with other data types (F32C4, U16C4); submitted draft PR: https://github.com/opencv/opencv/pull/26454.
- [OpenCV China] started working on a faster median filter: https://github.com/opencv/opencv/pull/26422. The goal is to replace the current "O(N)" and "O(1)" algorithms with the new "O(N)" algorithm which is significantly faster.
- [OpenCV China] continue working on replacing FLANN with Annoy. We found non-deterministic behaviour in Annoy, which has been also reported in the upstream repo; now trying to fix it.
- [OpenCV China] merged patch to 5.x that renames features2d to features, according to https://github.com/opencv/opencv/issues/24999
Vincent:
- TBD
Phil:
- OpenCV competition on bin picking (using robot arm); probably, could be a CVPR workshop, but also have ideas how to do it independently, as a regular OpenCV competition.
- PyCharm + OpenCV was presented last week with great success.
Shiqi:
- wechat QR detector and decoder. Students compared the new version by wechat team and the current "traditional CV" implementation.
- different implementations of aruco marker detection is in progress.
- libcamera backend for OpenCV is also in progress.
Alexander S (opencv.ai):
- Done:
  - Abduragim Shtanchaev:
    - Finishing refactoring of LSTM. The goal is to finally repair OCR models.
    - Fixed LSTM bidirectional for any batchsize #26391.
    - Some issue with test accuracy on TextDetection test. Working on it.
  - Alexander Smorkalov:
    - Prepated 4.x->5.x merge #26404, #26453
    - Added Universal Windows Package build to CI #26402, workflow #188.
    - Reviewed tested, helped to resolve issues ## 26364, 26396, 26418, 26394, 26405, 26414, 26403, 26379, 26401, 26400, 26397, 26390, 26254, 26383, 26384, 26391, 25624, 26446, 26444, 26443, 26420, 26434, 26405, 26155, 26437, 26379, 26435, 26414, 26348, 26451, contrib ## 3821, 3803, 3815, 3818, 3820, 3803.
    - Reworked KleidiCV integration #26303
    - Reworked FastCV integration to main #26316 and contrib #3811 repos. Pushed review changes to the own OpenCV fork branch FastcvHAL_1stPost.

2024-10-30

Vadim:
- [OpenCV China] found regressions in the new remap, some C3 and many C4 cases; found very good solution for C4; it's in progress now. C3 acceleration will probably be done a bit later.
- [OpenCV China] going to finalize and merge patch to OpenCV model zoo, which is the result of one of GSoC 2024 projects: block-wise quantization; many models have been significantly shrunk in size with very minor, if any, loss in accuracy: https://github.com/opencv/opencv_zoo/pull/270.
- [OpenCV China] the 1st Annoy patch has been submitted. Jia continues working on replacing FLANN with Annoy.
- [OpenCV China] patch with some fixes in the newly added intrin_math (v_sin, v_cos, v_exp etc.) to make it compatible with OpenCV 5 has been submitted and merged.
- [OpenCV China] started working on median filter acceleration. Preliminary results are very promising.
- Need to resume design activities & conversations with various teams about the new/extended HAL in OpenCV 5.
Shiqi:
- preparing program for F2F OpenCV core team meeting in 2024, December.
- the new QR detector & decoder; students still do evaluation, compare different libraries, models.
- libcamera support is being investigated.
- Aruco author suggested to try out aruco-nano, which is significantly faster than the current implementation of aruco in OpenCV.
Phil:
- preparing for F2F OpenCV meeting and on the conference in May.
- working with potential OpenCV partners.
- working on preparing OpenCV webinars.
- got links to some quite interesting OpenCV-based projects:
  - https://outline.georgs.lv/ - Make perfect cutouts for custom tool boxes using OpenCV in the browser
  - https://github.com/Sanjula-23/Hand-Gesture-Volume-Control-Code - This project provides a simple implementation of a hand gesture-based volume control system using Python, OpenCV, and MediaPipe. The program uses your webcam to detect hand gestures and adjust the system volume accordingly.
  - https://github.com/mikolajlubiak/terminal_animation Turn media (video/gif/image) into colored, animated ASCII art.
Vincent:
- tried to compile OpenCV 5 with the new dnn module; compiles fine.
- there is a big difference in 4.x version of calibration and 5.x. It would be nice to port some calibration code from 5.x back to 4.x.
Vladimir:
- working on improving UWP support in OpenCV 5
- OSoC (OpenCV summer of code) student has prepared the final report (on extending/improving RISC-V support in OpenCV, in particular FP16 branches); one of the final PRs is #26318.
Alexander S (opencv.ai):
- Done:
  - Merged TFLite & Caffe parsers for the new engine; TF parser is almost ready.
  - Merged some more refactored samples.
  - Merged a few patches from OpenCV China, see above
  - Abduragim Shtanchaev:
    - Almost done with refactoring the LSTM layer for the new graph engine. The new implementation will co-exist with the old one for now: the new one will be ONNX-compliant; the old one will support other backends. Working on fixing edge cases in test and correcting some tests
  - Alexander Lyulkov:
    - Works on TensorFlow parser for the new DNN engine.
  - Alexander Smorkalov:
    - Fixed and cleaned Caffe for the new dnn engine #26208.
    - Works on release documents.
    - Reviewed, tested, helped to resolve issues with ## 26379, 26316, 26287, 26326, 26364, 25708, 26348, 26384, 26374, 26330, 26336, contrib ##3811, 3815, 3800

2024-10-23

Vadim:
- [OpenCV China] remap acceleration patch has been submitted.
- [OpenCV China] Annoy integration is ready to be merged.
- Patch with the new engine merged. Block layout support is going to be added next.
Phil:
- talking to PyCharm team.
- preparing for OpenCV F2F in Dubai (Dec)
- preparing OpenCV conference in May in SJ.
Shiqi:
- have student for OpenCV development.
- special feature aruco-like detector may be submitted to OpenCV.
- the new QR code from Tencent is being tested now.
Alexander S (opencv.ai):
- Done:
  - Abduragim Shtanchaev
    - Works on refactoring LSTM layer for the new graph engine
  - Alexander Smorkalov:
    - Prepared 4.x->5.x merge #26358
    - Finalized Alex’s patch for the new engine support in TFLite.
    - Fixed Caffe specific tests with the new DNN engine in Alex’s patch.
    - Fixed KleidiCV error handling in build #26303. drop convertFp16 in favor of cv::Mat::convertTo #26327, contrib #3812.
    - Reviewed, tested, helped to resolve issues ## 26348, 26287, 26336, 26347, 26326, 26357, 26304, 26322, 26321, 26311, 26271, 26316
  - Alexander Lyulkov:
    - Works on TensorFlow parser for the new dnn engine.

2024-10-09

Vadim:
- [OpenCV China] accelerated warpAffine merged. warpPerspective acceleration patch is pending.
- [OpenCV China] correct univ. intrinsic-compliant implementation of v_exp, v_log and v_erf merged. v_cos & v_sin pending.
- 2 patches with new tests have been merged. One of them contains implementation of cv::norm() for CV_Bool.
Vincent:
- Tried to compile OpenCV 5.x. No major problems detected.
- Waiting for 4.11, because it's currently more important (because of smoother transition than to 5.0).
- std::vector issue needs to be solved before 5.0-release.
Shiqi:
- no big updates, since holidays in China have just finished.
- will meet with professor, AruCO author, this week.
- quite slow progress on integration of the new QR code detector from WeChat team; will meet with the student this week.
Satya:
- F2F meeting. Location is now defined (more or less). Date is ~mid of December.
Alexander S (opencv.ai):
- Finalizing several PRs before 5.0-alpha. 5.0-alpha is expected
- 5.0-beta and 5.0-release dates? The current plan is to squeeze both 5.0-beta and 5.0-release into 2024.
- Done:
  - Alexander Smorkalov
    - Fixed out-of-bound access bug in einsum layer with the new dnn engine #26272.
    - Fixed Yolo NAS test with the new DNN engine.
    - Extended multi-view calibration API, added bool mask support. Optimized intrinsics calibration in multi-view calibration pipeline #26248.
    - Relaxed conditions to allow 0d reshape for UMat. #26251. Fixed several OpenCL related failures after the new DNN engine integration.
    - Updated KleidiCV HAL to version 0.2. #26241
    - Reviewed, fixed helped to resolve issues with ## 25292, 26271, 25909, 26238, 26260, 25984, 26252, 26259, 26266, 26247, 26256, 26255, 26224, 26169, 26056, contrib ## 3709, 3801.
  - Abduragim Shtanchaev:
    - Proposed a solution for einsum layer out-of-bound access bug #26252
    - Fixed expand layer to support both new and old engines with static shape #26247.
    - Fix padding for new dnn engine #26220
  - Alexander Lyulkov
    - Fixed several bugs in the new dnn engine #26277 #26208
    - Working on modifying TFLite parser for the new dnn engine

2024-10-02

Vadim:
- the new engine is almost finished
Vincent:
- briefly tested 5.x (mostly compiling for now).
- One thing where OpenCV needs to be fixed (PR is about to be submitted) is CUDA modules compile errors when OpenCV is built without CUDA.
- OpenCV does not work correctly with vector<vector> (being wrapped into InputArrayOfArrays) in the case of some custom vector implementations.
Phil:
- finalizing competition about item picking algorithms.
- preparing for OpenCV F2F meeting in December. 2 locations are considered.
Gary:
- will hopefully have some big event in 2025, May. It will be a part of big conference. The main topic is 3D.
Shiqi:
- HEIF format in OpenCV: currently difficult to get it because of patent restrictions.
- Human body model: should be fine to use.
- Contacted to Aruco author; will have meeting with him next week.
Alexander S (opencv.ai):
- Done:
  - Abduragim Shtanchaev:
    - Working on a failing test caused by the new dnn engine integrated into the old one.
  - Alexander Smorkalov:
    - Collected 5.0-alpha scope: https://github.com/opencv/opencv/pulls?q=is%3Aopen+is%3Apr+milestone%3A5.0-alpha
    - Implemented calibrateMultiview tests and registerCameras test with real cameras data #26231.
    - Integrated registerCameras to mutliview calibration pipeline to support heterogeneous setups (pinhole + fisheye). Added test with real camera data. #26213
    - Reworked multiview calibration API and updated documentation #26221
  - Fixed gemm layer regression in the new DNN engine branch.
    - Reviewed, tested, helped to resolve issues with ## 26230, 26229, 26228, 26227, 26225, 26222, 26220, 26189, 26210
  - Alexander Lyulkov:
    - Modified Caffe parser for the new dnn engine: https://github.com/opencv/opencv/pull/26208
    - Fixed incorrect argument handling in forward functions for the new dnn engine
    - Working on fixing tests, which were found after enabling Caffe parser for the new model

2024-09-11

Vadim:
- Continue working on the new DNN engine, several bugs have been fixed and ~10 utility layers (Concat, Reshape, Shape, Slice, Split, Squeeze, Transpose, Unsqueeze, ...) have been reimplemented/added to reduce number of test failures.
- [OpenCV China] work on warpAffine is at final stage. All tests pass.
- 2 more refactored samples by Gursimar Singh have been finally merged, 1 is to be merged within a few days, 1 more is to be reviewed.
Vincent:
- GSoC PR on WebP animation + new API is still in progress, but is in good shape already.
- Discovered some problems with OpenCV JS bindings. Going to submit fixes soon. It looks like OpenCV.js is not used actively now, because some of the bugs are rather obvious (e.g. not good support for nested namespaces).
Phil:
- continue working on store at OpenCV.org.
- preparing for 1-day conference in San Jose.
- got 1 more OpenCV.org silver member, and talking to some other potential members.
Shiqi:
- GSoC project on GIF codec: the student finished PR; it's ready for review. He will add some more text & description.
- Another student is working on the new QR detector & decoder (based on the new code drop from WeChat team).
- Continue exploring options for the next OpenCV core team F2F meeting next year. Talked to some guys in Hongkong; other options are China (Shenzhen) and Vietnam.
Alexander S (opencv.ai):
- Done:
  - Abduragim Shtanchaev:
    - Working on fixing layer ConvTranspose for new graph engine.
  - Alexander Smorkalov:
    - Prepared 4.x->5.x merge #26135
    - Added Leaky RELU support for TFLite #26132
    - Excluded nullptr leak to arithmetic HAL got from empty Mat #26125
    - Added HAL interface for Lukas-Kanade optical flow #26143
    - Reworked zlib-ng update to 2.2.1 #26113. Fixed zlib-ng build issue zlib-ng/zlib-ng#1774
    - Reviewed, tested, helped to resolve issues with ## 26134, 26124, 26120, 26119, 26117, 26116, 26115, 26114, 26110, 26106, 26042, 26025, 25326, 26101, 25984, 26012, contrib ## 3490, 3791
  - Alexander Lyulkov:
    - Added 2x-5x faster version of blobFromImage and blobFromImages in dnn #26127
  - Rostislav Vasilikhin:
    - More types added to mixed arithmetics tests: #26134

2024-09-04

Vadim:
- [OpenCV China] Continue to work on warpAffine; fixed a few regressions.
- [OpenCV China] Continue work on cleaning up univ. intrinsics and the new math functions implementations.
- Draft new DNN engine has been submitted: https://github.com/opencv/opencv/pull/26056.
- Need to make a video out of complete GSoC projects.
Vincent:
- Reviewed PRs submitted by GSoC student
Gary:
- Talked with GSoC student about the final stage of the project (since this is long-duration GSoC project). She is working on some real demo of those algorithms (dust3r and mast3r).
Shiqi:
- GSoC student did some comparison with other libraries (like libgif). Performance of his GIF codec has been improved, but PR is not updated yet.
Phil:
- Preparing for OpenCV store launch.
- Further negotiations with current and possible future sponsors. Added some of them to OpenCV Slack channel.
- Looking for more OpenCV episodes (video talks) for OpenCV site.
Alexander S (opencv.ai):
- Done:
  - GSoC 2024: trying to finalize the project on GTK4 + OpenGL support in OpenCV (#25958). Most likely a separate branch of code is needed to handle {GTK3 and GTK4} + OpenGL properly.
  - Some improvements in DNN have been done by Abdurraheem, such as added HardMax and GatherND layers etc.
  - Fixed build issues with CUDA 12.6 #3785; there was problem with std::tuple.
  - Got rid of CAROTENE_NEON_ARCH and use standard __ARM_ARCH check #26091
  - Reviewed, tested, helped to resolve issues with ## 26079, 26105, 26071, 26102, 26093, 26023, 26094, 26106, 26090, 26042, contrib ## 3786, 3784, 3782, 3778
  - Alexander Lyulkov can take part in OpenCV part time for now. He is going to submit faster blobFromImage.
  - Did some research on how to improve OpenCV efficiency in the case of 32-bit mode on ARMv8 platforms; there are some customers who may need it.

2024-08-28

Vadim:
- Continue working on the new DNN engine: https://github.com/opencv/opencv/pull/26056
- [OpenCV China] Extended the new warpAffine implementation to support U8/U16/FP32, C1/C3 inputs: https://github.com/opencv/opencv/pull/25984. FP16 should be easy to add. Currently working on adding C4 support. Discovered a few regressions in accuracy and speed. Trying to address them as well.
- [OpenCV China] GSoC 2024 project on block-wise quantization yielded some great results: https://github.com/opencv/opencv_zoo/pull/270. We observed no regressions in the terms of inference speed.
- [OpenCV China] refactoring of universal intrinsics to add math functions in a proper way and to get rid of overloaded operators (in favor of v_add, v_sub, v_or etc.) is in progress.
- Discussed GSoC project status: 3 projects are successful, 2 projects are not and on the remaining 2 we'll decide this week.
Gary:
- ok to continue OpenCV summer of code (on enhanced RISC-V RVV support)
Phil:
- trying to extend websites with more options to donate to opencv.org. Also, working on expanding opencv store (that will offer not just mugs or t-shirts, but also products from partners, such as depth cameras etc.)
Shiqi:
- student has finished work on GIF codec.
- probably should update OpenCV GSoC ideas list (maybe create GSoC 2025 in advance). Specifically mark topics which ended up with successful projects (e.g. put links to PRs) to avoid repetition of things that are already done.
Alexander S (opencv.ai):
- Q: Android @ RISC-V support status? Shall OpenCV get such configurations in CI? A: no concrete information as for now.
- OpenCV + Unity: some potentially interesting projects & competitions are possible.
- Done:
  - Abduragim Shtanchaev:
    - Finished on issue with Einsum layer 1d falling NEON backend #26059
    - Working on implementing the Hardmax layer #26079.
  - Alexander Smorkalov:
    - Prepared 4.x->5.x merge #26073
    - Reviewed and benchmarked new warpAffine #25984.
    - Updated Android SDK build environment to NDK 27, Android build tools 34 and Gralde 8.10. Added 16kb pages support #26057. IPP update needed #26072.
    - Fixed Python and Java bindings in new DNN engine #26056
    - Fixed minMaxIdx indexing issue with HAL #26080
    - Reviewed, tested, helped to resolve issues ## 26053, 26025, 26012, 26075, 26022, 26069, 26063, 26062, 26061, 26059, 26054, contrib ## 3774, 3782, 3778
  - Alexander Lyulkov:
    - Working on adding tensor core tf32 convolutions to CUDA dnn backend

2024-08-21

Vadim: ...
Shiqi:
- GIF encoder and decoder are ready for review (PR #25691); there are regression tests. encoder of true-color images is also ready, it uses some color quantization.
Alexander S (opencv.ai):
- GSoC:
  - Initial support for GTK3 + OpenGL in HighGUI is enabled. Support for GTK4 is just started. Also, there are some conflicts with the current OpenGL support in core module.
- others:
  - Abduragim Shtanchaev:
    - Finished Python boolean bindings issue. #26026
    - Figured issue where CUDA would fall back on CPU #26021
    - Working on issue with Einsum layer 1d falling NEON backend
  - Alexander Smorkalov:
    - Disabled quantize-dequantize tests with CUDA after 4.x->5.x merge #26050.
    - Fixed memory leak in dshow camera backend #26045.
    - Finished JS support for opencv_contrib #25986.
    - Reviewed, tested, helped to resolve issues ## 25950, 26052, 25968, 26043, 26048, 26009, 26026, 26046, 26038, 26034, 26041, 25980, 26019
    - Working on adding support for bool type into various places in OpenCV, e.g. to be able to use boolean mask (CV_Bool) in every place where CV_8U mask is now expected.
  - Alexander Lyulkov:
    - Added integer types to OpenCL dnn backend #26048
  - Rostislav Vasilikhin:
    - Fixed rounding mode for test: #26052

2024-08-14

Vadim:
- [OpenCV China] there is some progress with warpAffine optimization; more types (F32, F16) and more channel options (C=1, C=4) will be added. warpPerspective will be done in a separate PR. It looks like double-precision division is needed to compute coordinates accurately enough in the case of warpPerspective.
- [OpenCV China] GSoC project on block-wise quantization goes well. The first PR has been merged. Now the student will concentrate on fixing ONNX loader to handle blockwise-quantized models efficiently (move all dequantization steps outside of the inference loop)
- Wanli finished work on v_exp, v_log, v_erf; now finalizing work on v_sin, v_cos, v_sincos.
- GSoC: just ~2 weeks left before "soft" code freeze, 3 weeks before "hard" code freeze.
Shiqi:
- GSoC student mostly finished GIF encoder; decoder was finished earlier. Both parts are ready for review.
- Presented information about OpenCV, mostly DNN module, at the conference in Vietnam. ~100 participants, ~50% of them use OpenCV.
Gary:
- GSoC students should prepare docker image with the 3D reconstruction framework/demo and the respective demonstration video.

2024-08-07

Shiqi Yu:
- GIF decoder is finished. PR is modified, including the new tests.
- Most parts of GIF encoder are finished as well
- Will attend a conference in China next week: http://bic-conference.org/2024/invited-speaker.php. Will give a talk on OpenCV and its use for CV on the edge.
Phil:
- working on 'OpenCV store' at opencv.org.
- initial preparations for OpenCV 5 webinars
Alexander S (opencv.ai):
- Done:
  - Abduragim Shtanchaev:
    - Still working on reproducing the accuracy issue in ConvLayer with padding.
    - Working on issue where layer fall back to CPU (when inference is made with CUDA) backend #25985, #25982
  - Alexander Smorkalov:
    - Prepared 4.x->5.x merge #25998.
    - Fixed cv::Mat::allocator bug, added tests #25979
    - Improved error handling in image codecs #25975
    - Finalized algorithm hint support for cvtColor #25932
    - Finished distortPoints variant for fisheye #25943
    - Split Javascript white-list to support contrib modules #25986
    - Fixed asserts in fillPoly #25996
    - Work on for Mask support with CV_Bool in remaining modules #25951 is still in progress. ORB and some other algorithms are sensitive to the “true” value.
    - Reviewed, tested, helped to resolve issues with ## 25991, 25987, 25970, 25978, 25972, 25786, 25984, 26009
  - Alexander Lyulkov:
    - Unified build.gradle files into one template #26009
    - Starts working on shape inference and new types in OpenCL backend in DNN (5.x).
  - Rostislav Vasilikhin:
    - Analyzed pyrDown issues:
      - IPP code is broken, opened an issue #25976 (IPP code is removed now)
      - Is not consistent between ROIs, filed an issue #25990
      - Extended pyrDown HAL to handle BORDER_ISOLATED correctly #25970.

2024-07-31

Abduragim Shtanchaev:
- Finalized on sample for Diffusion model. Opened PR
- Found out that Conv2D implementation in opencv has slightly deviating results (compared to torch and onnxruntime), which results in slightly different results in Diffusion model. Opened PR
Alexander Smorkalov:
- Implemented CV_Bool support as mask in remnaining modules #2595. Work on ORB issue #25974
- Added fisheye::distortPoints with non-identity projection matrix #25943
- Prepared 4.x->5.x merge #25953
- Finalized Algorithm hint option in cvtColor #25932. Added several more cvtColor cases for YUV 25949.
- Added GTK-4 to CI (GHA #178).
- Reviewed, tested, helped to resolve issues with ## 25644, 25886, 25958, 22551, 25950, 25972, 25965, 25961, 25955, 25952, 25944, 25970, contrib #3771, 3600, 3607
Alexander Lyulkov:
- Prepared reproducer for #25886
- Works on #24686
Phil
- Donation portal set up w/Goole Microsoft, Meta
- Generative AI for that
ARM, Qualcom optimizations ...
Gary
- Set up a convo on Spatial Vision

2024-07-24

Vadim:
- [OpenCV China] started working on accelerated warpAffine (~2.5x acceleration is expected). warpPerspetive and remap will follow.
- [OpenCV China] features2d patch finally passes all tests. Will me merged soon.
- [OpenCV China] flann=>annoy replacement patch is ready for review.
- Experimental inference engine is running some simple image classification models. Now need to port the engine to the existing opencv dnn.
Vincent:
- will be on vacation in August.
- doing some internal testing of OpenCV 5.0-pre. it's more or less good, 1D Mat introduction as well as the new text rendering engine yield some test failures.
Phil:
- PyCharm + OpenCV = friendship!
- A few exciting OpenCV events are upcoming
Gary:
- Doing some spatial computing experiments
Alexander S (opencv.ai):
- Done:
  - Abduragim Shtanchaev:
    - Working on sample for Diffusion model
    - Run first diffusion sample in opencv
    - Working on figuring why torch vs opencv results do not match exactly
  - Alexander Smorkalov:
    - Extended tests for CV_Bool support as mask #25902
    - Extended cvtColor with implementation hint and added corresponding API #25932. Works on new cvtColorTwoPlanes to support BGR -> NV21/NV21 conversions.
    - Added fisheye::distortPoints with non-identity projection matrix #25943
    - Reviewed, tested, helped to resolve issues with ## 25807, 25942, 25941, 25938, 25936, 25934, 25824, 25874, 25883, contrib #3607, 3600

2024-07-17

Vadim:
- [OpenCV China] GSoC project on block-wise quantization. Tool is more or less ready and tested on Resnet50 successfully.
- [OpenCV China] Features2d refactoring is almost finished.
- [OpenCV China] Several activation functions have been optimized.
Vincent:
- no updates
Gary: *
Satya:
- new gold member joined opencv.org. another one is on the way.
- working on the next OpenCV competition.
Shiqi:
- Met with S.... company. They want to collaborate with OpenCV, donate h/w to CI.
- GSoC: the student continues working on GIF decoder, finished alpha support.
Phil:
- opencv.org is available
Alexander S (opencv.ai):
- Done:
  - Abduragim Shtanchaev:
    - Fixed GPT2 sample. #25868
    - Fixed reduce layer to correctly support keepdims parameter #25101
    - Fixed reduce layer parser to support inputs as initializers #25917
    - Working on inpainting example using Diffusion model
  - Alexander Smorkalov:
    - Prepared 4.x->5.x merge (#25915); should be merged today (17th July)
    - GSoC student finished patch for OpenGL+GTK3/GTK4 support in OpenCV HighGUI.
    - Added CV_Bool support as mask to TS and Core modules (#25902)
    - Fixed several issues found after PRs merge ## 25912, 25911
    - Reviewed, tested, helped to resolve issues ## 25913, 25822, 25824, 25903, 25900, 24941, 25899, 25630, 25252, 25880, 25147, 25894, 25808, 25864, contrib ## 3731, 3753, 3748, 3607, 3619

2024-07-10

Vadim:
- [OpenCV China] v_exp, v_log & v_erf was merged; acceleration of various activations (GELU, Mish, Swish ...) is in progress
- [OpenCV China] several PRs have been merged. Almost finalized patches with features2d cleanup and flann => annoy replacement.
- GSoC: mid-term evaluations are going on.
Vincent:
- png encoder fixes.
- prototype for "const Mat". C++ 11 is required, C++ 20 is nice (https://github.com/opencv/opencv/pull/25869).
Gary:
- looking at/playing with Dust3r and Mast3r algorithms for dense 3d reconstruction.
Shiqi:
- GSoC student is finalizing decoder and started working on encoder.
Phil:
- Making some nice changes to opencv.org site.
Alexander S (opencv.ai):
- Done:
  - Abduragim Shtanchaev:
    - Create a sample for inferencing GPT2 model. #25868
    - Working on sample for Diffusion model.
  - Alexander Smorkalov:
    - Finished work on Algorithm implementation hints API. PR #25792 for GaussianBlur.
    - Helped to fix issues with CUDA_FP16 #25880.
    - Helped to finalize PR “Add a new function that approximates the polygon bounding a convex hull with a certain number of sides” #25607
    - Added CI pipeline with Ubuntu 24.04 for 5.x (#175)
    - Update list of supported Python versions in Windows Mega Pack (#173)
    - Reviewed, tested, helped to resolve issues ## 24941, 25147, 25885, 25808, 25864, 25872, 25844, 25877, 25838, 25888, 25252, 25886, 25874, Contrib ## 3753, 3748.
  - Alexander Lyulkov:
    - Fixed Kotlin plugin search in Android projects #25856
    - Reproduced ORB issue on Android #24816

2024-07-03

Vadim:
- [OpenCV China] v_exp & v_log merged
- contribution from Mu Zihao: imread() now has 'RGB' flag
Vincent:
- patch for jpeg decoder bug was submitted: https://github.com/opencv/opencv/pull/25864
Phil:
- various OpenCV activities, attracting new sponsors
- preparing for the new OpenCV competition
Gary:
- talked to Kornia team, they are going to convert some of their code to Rust.
Alexander S (opencv.ai):
- Probably will release minor OpenCV (Android only) with extra ARM optimizations soon.
- Done:
  - Rostislav Vasilikhin:
    - Prepared HAL for setTo() (draft PR #25819).
  - Alexander Smorkalov:
    - Prepared 4.x - > 5.x merge (#25840)
    - Enabled Ubuntu 24.04 build for 5.x (#25855). Added pipeline, fixed warnings (#25848).
    - Fixed JS bindings test in 5.x (#25863)
    - Fixed 5.x build for Linux 32-bit (#25860)
    - Added lut support for all new types in 5.x (#25847)
    - Fixed CuDNN runtime version check for CuDNN 9+ (#25841).
    - Reviewed, tested, helped to resolve issues with ## 25813, 25836, 25838, 25820, 24941, 25800, 25756, 25857, 25799, 25781, 25809, 25630, 25794, 25410, 25827, 25826
  - Alexander Panov:
    - Seek leave.
  - Alexander Lyulkov:
    - Fixed Kotlin dependency in OpenCV4SDK #25856.
  - Abduragim Shtanchaev:
    - Finished adding support for Batched Attention from Pytorch. #25851

2024-06-26

Vadim:
- [OpenCV China] submitted PR with flann/miniflann replaced with Spotify's Annoy-based ann engine.
- [OpenCV China] added new depthToSpace layer into DNN
- [OpenCV China] finalized v_exp implementation
- keep working on dnn engine
- reviewing samples by Gursimar
Satya:
- 2 new gold members joined opencv.org
- plan to have annual OpenCV dev forum, need to choose the proper place
Shiqi:
- GSoC: students resumes work on animated gif project after exams.
- attended CVPR. Great event: almost 10000 participants.
Gary: *
Vincent:
- GSoC: going to connect the two students working on animated images support to use the same API
Alexander S (opencv.ai):
- Done:
  - Rostislav Vasilikhin:
    - Vacation 1 week
    - Erode/Dilate tests: test data is actually compared to reference data (complicated test code) equalizeHist() accuracy tests: PR is done & merged #25759
    - HAL for setTo(): in progress
  - Alexander Smorkalov:
    - Reviewed and tested 4.x->5.x merge by Maksim in #25745. The check was done by own independent merge and compare. Proposed solutions for new HAL integrations for projectPoints and minMaxIdx.
    - Made experiments with remaining Carote functions that was not used or disabled before (#25799, #25801). The last one has accuracy issues, closed.
    - Proposed ImplementationHint option for popular OpenCV functions. Prepared implementation on top of IPP with tests for GaussianBlur #25792.
    - Works on more accurate replacement of #24927.
    - Reviewed, tested, helped to resolve issues with ## 25809, 25410, 25569, 25811, 25779, 25775, 25813, 25805, 25791
  - Alexander Panov:
    - Works on new implementation of findCircles grid for asymmetric/symmetric case (issue #24971)
    - Testing algorithm for searching for board lines
  - Abduragim Shtanchaev:
    - Added support for yolov9 and v10 in opencv. #25794
    - Works on Attention support in 5.x and makes experiments with running GPT2 in OpenCV.

2024-06-19

Vadim:
- [OpenCV China] prepared PR with depthToSpace DNN operator
- [OpenCV China] added v_log intrinsic; fixed v_exp implementation
Gary: *
Vincent:
- Alternative to 'universal intrinsics': https://github.com/google/highway
- GSoC project: Suleyman is working on WebP and Png animations.
- Prepared document with description of image animation API - how it should be done in OpenCV: https://docs.google.com/document/d/1_9_tLaMhJZ-tbqmPPrBgwiS0Ii4GAZWH0z6-wtqfefg/edit#heading=h.1l5qhfbcgwxs
- Inter area resize optimization: resumed work on it.
Alexander S (opencv.ai):
- Rostislav Vasilikhin:
  - PRs with --hwasan flag added: #25718, #25746 .
  - Instruction written Checked all main and most contrib tests, filed found HWASan issues: #25735 #25741
  - Filed other failed G-API tests: #25748 #25749
  - Added accuracy tests for equalizeHist() #25759.
- Alexander Smorkalov:
  - Prepared OpenCV-Python 4.10 release with NumPy 2.0 support (Python #1001, #1002).
  - Fixed legal issue in calibration pattern generator (svgfig.py lib) #25778
  - Implemented LUT for FP16, added accuracy tests #25787
  - Extended bilateralFilter tests #25788
  - Fixed 2 issues found by HWASAN: #25746, #25780
  - Fixed result buffer overflow in intersectConvexConvex_ for non-convex input #25725
  - Reviewed, tested, helped to resolve issues with ## 25519, 25630, 25770, 25410, 25746, 25763, 25771, 25569, 25768
- Alexander Panov:
  - Works on new implementation of findCircles grid for asymmetric/symmetric case (issue #24971)
  - Added an algorithm for finding the borders and corners of the board.
  - Boards with large distortions have been tested.
  - An algorithm for searching for board lines is being added now.
- Alexander Lyulkov:
  - Fixed old OpenVINO versions support in dnn #25739
  - Added fix to Android H264/H265 Encoding, that was proposed by the community #25716
  - Added more types support for dnn layers #25755
- Abduragim Shtanchaev](mailto:abduragim.shtanchaev@xperience.ai):
  - Working on adding support for yolov10 in opencv

2024-06-05

Gary:
- keep working on GSoC 2024 projects regarding 3D stuff.
Vincent:
- OpenCV 4.10.0 has been briefly tested, so far so good.
- Pinging regarding issue in opencv.js (https://github.com/opencv/opencv/issues/25239)
- Continue to work with Suleyman on WebP (GSoC 2024).
Alexander S (opencv.ai):
- Done:
  - Rostislav Vasilikhin:
    - HAL for mul8x16x: discussed how to fix it, fixed it, merged
    - HAL for sub8x32f: fixed according to what’s discussed, merged
    - HAL for projectPoints(): fixed according to discussion, merged
    - fcvAddScalarf32 HAL integration: in progress
    - YUV cvtColor:
      - Investigated all YUV conversions, color conversion codes, coefficients and HAL availability options
      - docs PR prepared: #25616
  - Alexander Smorkalov:
    - Prepared KleidiCV update to version 0.1.0 (#25618)
    - Prepared libjpeg-turbo update to version 3.0.3 (#25623)
    - Updated OpenJPEG to version 2.5.2 (#25611)
    - Fixed build issue with CuDNN 9 (#25412)
    - Updated dependencies for OpenCV-Python build (#984)
    - Works on 4.10.0 release (#25615)
    - Reviewed, tested, helped to resolve issues ## 25116, 25595, 25580, 25617, 25612, 25511, 25601, 25582, 24804, 25574, 25506, 24903, 25583, contrib ## 3627, 3623, 3726, python #981.
  - Alexander Lyulkov:
    - Added bool support to DNN 5.x #25605
    - Checked issues #25537 and #21519. Newest OpenCV 4.x already works fine
    - Works on CuDNN 9 LSTM issue (#25412)
  - Abduragim Shtanchaev:
    - Finalized issue related to Einsum. PR
    - Working on model import/read issue 4.x
  - Alexander Panov:
    - back to OpenCV team; will be working on improving circle grid detector.

2024-05-22

Had discussion about integration of WeChat compatible barcode detector
- Embedded people want seperable independent modules
Need to start reporting on GSoC progress

Rostislav Vasilikhin:
1. HAL for mul8x16x: discussed how to fix it, fixed it, merged
2. HAL for sub8x32f: fixed according to what’s discussed, merged
3. HAL for projectPoints(): fixed according to discussion, merged
4. fcvAddScalarf32 HAL integration: in progress
5. YUV cvtColor:
  1. Investigated all YUV conversions, color conversion codes, coefficients and HAL availability options
  2. docs PR prepared: #25616
Alexander Smorkalov:
1. Prepared KleidiCV update to version 0.1.0 (#25618)
2. Prepared libjpeg-turbo update to version 3.0.3 (#25623)
3. Updated OpenJPEG to version 2.5.2 (#25611)
4. Fixed build issue with CuDNN 9 (#25412)
5. Updated dependencies for OpenCV-Python build (#984)
6. Works on 4.10.0 release (#25615)
7. Reviewed, tested, helped to resolve issues ## 25116, 25595, 25580, 25617, 25612, 25511, 25601, 25582, 24804, 25574, 25506, 24903, 25583, contrib ## 3627, 3623, 3726, python #981.
Alexander Lyulkov:
1. Added bool support to DNN 5.x #25605
2. Checked issues #25537 and #21519. Newest OpenCV 4.x already works fine
3. Works on CuDNN 9 LSTM issue (#25412)
Abduragim Shtanchaev:
1. Finalized issue related to Einsum. PR
2. Working on model import/read issue 4.x

To Dos

Name
- todo

2024-05-15

Vadim:
- [OpenCV China] submitted experimental CLBlast support that tries to load dynamic CLBlast library and use it. We decided to repair it more (fix accuracy and low perf issues on Mac M-chips and Intel iGPU).
- [OpenCV China] did more benchmarks opencv flann vs flann vs annoy vs brute force (including nanoflann). It looks like in high-dimensional spaces Annoy is the fastest.
Vincent:
- contacted to Suleyman, discussed GSoC 2024 goals and execution plan.
Shiqi:
- GSoC student: GIF decoder is done, works well, will submit PR soon. GIF encoder will be the next task.
- The new Wechat QR detection/decoding pipeline is in progress, OpenCV-compatible version will likely be ready within 1 month.
Phil:
- CVPR 2024 is comming (2nd half of June). Embedded Vision Summit next week. Preparing materials for the events.
Gary:
- Got back from Korea. Started talking to GSoC students.
Alexander S (opencv.ai):
- Done:
  - Rostislav Vasilikhin:
    - Added HAL for equalizeHist() #25565
    - Extended HAL interface for sub operation, fixed review notes.
    - Investigated extra HAL cases for cvtColor.
  - Alexander Smorkalov:
    - Works on 4.x->5.x merge
    - Added Ubuntu 24.04 to CI fixed warnings and test failures (contrib #3734, CI #172).
    - Worked on Wayland support. Reviewed, tested, helped to resolve issues with ## 25561, 25551, 25590, 25510. Works on Wayland support via Qt in OpenCV Python #784
    - Reviewed, tested, helped to resolve issues with ## 25583, 25577, 25576, 25567, 25566, 25565, 25563, 25559, 25558, 25554, 25582, 25487, 25390, 25506, 25580, 25555, 25503, 25592, 25515, 25326, 25326
  - Alexander Lyulkov:
    - Added int8 and uint8 input support to CUDA in dnn 5.x #25555
    - Working on adding bool support to dnn 5.x
  - Abduragim Shtanchaev:
    - Fixed issues related to types in Arg layer
    - Added support for 0D vector in Einsum (still add some corner cases)

2024-05-15

Alexander S (opencv.ai):
- Done:
  - Rostislav Vasilikhin:
    - HAL for LUT added: PR #25554
    - HAL for mul8x16x: trying to make it properly
    - HAL for Otsu threshold: fixed, merged
    - HAL for projectPoints(): fixed by code review comments
    - HAL for integrate() and imageDiff(): investigated how to make them
  - Alexander Smorkalov:
    - Added KleidiCV as HAL implementation (#25443)
    - Fixed several HAL related bugs (## 25563, 25553, 25558, 25532).
    - Works in Ubuntu 24.04 support (CI, build warning fixes, test failures).
    - Reviewed, tested, helped to resolve issues with ## 25487, 25557, 25551, 25554, 25506, 25548, 25484, 25196, 25525, 24804, 25542
  - Alexander Lyulkov:
    - Added int8 and uint8 input support to dnn 5.x #25555
  - Abduragim Shtanchaev:
    - Finishing up last open PRs related to 0/1D inputs support. Layers such as tile, arg, gather, einsum

2024-04-24

Vadim:
- [OpenCV China] working on adding a better FLANN substitution to OpenCV (Annoy perhaps?)
- [OpenCV China] preparing PR with CLblast-based GEMM optimization
- [OpenCV China] continue to work
- pre-ranked GSoC 2024 proposals, assigned potential mentors. Waiting for the final slot allocation.
Vincent:
- There is OpenCV-Python+numpy related bug. Will provide fix.
Dmitry (Intel):
- Scheduled meeting about OpenCV 5.0 HAL API with IPP team.
Shiqi:
- Could possibly help with demos for CVPR 2024.
- WeChat's new QR detector & decoder: all 3 models (detector, segmentation, 4-corner detector) work smoothly. And they are not big: QR detector - 2.99MB (ONNX), segmentation - 754KB (ONNX), corner detector - ??? (to be reported later).
  - Alexander: need to compare WeChat solution with our current solution: https://github.com/opencv/opencv_benchmarks/tree/develop/python_benchmarks/qr_codes.
Phil:
- OpenCV is participating in Embedded Vision Summit 2024: https://embeddedvisionsummit.com. Going there together with Satya.
- Preparing for CVPR 2024 as well. Need to think about demos.
Alexander (opencv.ai):
- Many improvements in DNN in 5.0. ONNX compliance rate increased from 22% to 29%.
- Rostislav Vasilikhin:
  - Implemented performance test for cv::solve and cv::SVD
  - Fix build for Windows: bad var names in testdata generator #25451
- Alexander Smorkalov:
  - Prepared 4.x-> 5.x merge (#25460). The new one is coming.
  - Added JS bindings for DNN based face detector (#25430).
  - Added meanStdDev HAL API entry (#25483).
  - Ported Windows ARM64 CI to 5.x, fixed warnings (#25428).
  - Made several GSoC student interviews.
  - Reviewed, fixed, helped to resolve issues with ## 25433, 25324, 25465, 25467, 25468, 24903, 25435, 25453, 25450, 25421, 25469, 25480, 25473, 25463, 25427, 25411
- Alexander Lyulkov:
  - Works on int/int64 support with OpenVINO DNN backend in 5.x (#25458).
- Abduragim Shtanchaev:
  - Updated ONNX conformance filters (#25465, #25480, 25442). Triaged remaining test failures.
  - Merged 0d/1d support in BatchNorm (#25420). Works on review remarks.

2024-04-17

Vadim:
- [OpenCV China] experiments with CLBlast go on. We found that it's the fastest available OpenCL implementation of MatMul, but it's ~10x slower than Cublas' gemm on NVidia (which probably uses specialized hardware for matrix multiplication)
- [OpenCV China] cleaning up Features2D module (renamed to Features in 5.x), samples and DNN importers is going on.
- GSoC 2024: now reviewing all received proposals.
Shiqi:
- WeChat team provided updated models (in .onnx format) for detection and superresolution of QR codes.
Phil:
- negotiations with JB, as well as other possible OpenCV sponsors, go on.
- preparations for Embedded Vision Summit are being made
Vincent:
- fixed possible bugs in ffmpeg wrapper, working on updates for WebP wrapper.
Alexander S (opencv.ai):
- Done:
  - Rostislav Vasilikhin:
    - 3D model benchmark: finished except minor code review comments left
    - Finished rendering benchmark: cmd args, help, color fixes, temp files, stat output, generate normals, speedups
    - Got statistics (looks good)
    - Test data generator rewritten to support user models
    - Various fixes for model saving/loading: texCoords, mixChannels #25382 merged
    - SVD and Solve perf tests: in progress
  - Abduragim Shtanchaev:
    - Working on 0/1D support to layers. Almost finished, ScatterND is left.
  - Alexander Panov:
    - Merged PRs Move Charuco/Calib tutorials and samples to main repo #25378, #3708
    - reviewed Take first suitable neighbor in ChessBoardDetector::findQuadNeighbors #25365
    - Work on asymmetric findCirclesGrid - findCirclesGrid can't detect some asymmetric grids #24971
    - drawing and researching detected results
    - found problems with non-square patterns
    - testing CALIB_CB_CLUSTERING method
  - Alexander Smorkalov:
    - Prepared 4.x->5.x merge (#25422)
    - Extended HAL API for GaussianBlur and remap.
    - Fixed jpeg-turbo diagnostics and build options in default configuration (#25388).
    - Added CI pipeline for Windows ARM in 4.x. And 5.x. Work on warnings is in progress (#25383, #25428).
    - Reviewed, tested, helped to resolve issues with ## 25415, 25413, 25407, 25406, 25402, 25424, 25304, 25382, 25378, 25431, 25432, 25387, 25351, contrib ## 3722, 3719, 3708, 3717.
  - Alexander Lyulkov:
    - Fixed range layer and added tests PR PR
    - Working on modifying OpenVINO dnn backend for the new dnn with different types

2024-04-10

Vadim:
- [OpenCV China] compared speed of OpenCL-based gemm implementation in different libraries. So far, CLblast is the fastest, but we did not compare it against Intel-specific code.
- [OpenCV China] continue to cleanup OpenCV 5.x code, i.e. objdetect and features2d modules, as well as samples.
- GSoC 2024: made a spreadsheet of received proposals, have time till April 22nd to review them.
Vincent:
- sjpeg is the recommended solution to produce high-quality & small-size jpegs: https://github.com/webmproject/sjpeg
Shiqi:
- Had meeting with Wechat team. They are going to submit new models (in .onnx format) for qr detection, corner localization, super-resolution.
Phil:
- ...
Alexander S (opencv.ai):
- Done:
  - Rostislav Vasilikhin:
    - 3D model benchmark:
    - Texture coordinates fixed: #25382
    - Texture is now loaded correctly, finishing rendering benchmark is in progress
  - Abduragim Shtanchaev:
    - ...
  - Alexander Panov:
    - Prepared PR Move Charuco/Calib tutorials and samples to main repo #25378
      - update/fix charuco_detection.markdown, images and samples
      - update/fix charuco_diamond_detection.markdown, images and samples
      - update/fix aruco_calibration.markdown, images and samples
      - update/fix aruco_faq.markdown and images
      - moving tutorials, samples and tests to main repo
    - Prepared and merged PR add detectDiamonds to diamond test #3713
    - Work on asymmetric findCirclesGrid - findCirclesGrid can't detect some asymmetric grids [#24971(https://github.com/opencv/opencv/issues/24971)
  - Alexander Smorkalov:
    - Prepared 4.x->5.x merge (#25384), backport changes in new findContours algorithm to 4.x (#25385)
    - Introduced Windows ARM64 builds in CI for 4.x branch, suppressed and fixed warnings (##25383, GHA #155, 25375).
    - Introduced new HAL entries for BilateralFilter (#25343), Transpose2d (#25342), works on GaussianBlur.
    - Reviewed, tested, helped to resolve issues with ## 25381, 25380, 25379, 25372, 25371, 25369, 25368, 25364, 25357, 25356, 25339, 25335, 25333, 25330, 25328, 25327, 25325, 25322, 25387, 25351, 25349, 25317, 25304.
  - Alexander Lyulkov:
    - Added tests for adaptiveThreshold and sepFilter2D #25371 #1167 Working on modifying OpenVINO dnn backend for the new dnn with different types

2024-04-03

Vadim:
- [OpenCV China] working on cleaning up OpenCV 5.x
- [OpenCV China] working on accelerating transformers (4.x & 5.x)
- [OpenCV China] working on the new inference engine
- GSoC 2024: received 100+ proposals. Have ~2.5 weeks to review them.
Vincent:
- ...
Shiqi:
- Wechat team are going to contribute new/improved models to OpenCV
Phil:
- ...
Alexander S (opencv.ai):
- Done:
  - Rostislav Vasilikhin:
    - 3D model benchmark:
    - mesh loading/saving: Vec3i for indices supported, [0, 1] floats for colors instead of uchar for better accuracy
    - tool was moved to separate PR: #13@benchmark (draft)
    - Found & fixed a bug in PLY loader: wrong offset processing ruined the least byte of x coordinate which led to ruined precision
    - fixing output argument processing appeared after python code is checked; adding python test is in progress
    - fixed testdata generator to support arbitrary camera parameters
  - Abduragim Shtanchaev:
    - Finishing all open PRs related to 0/1D support
    - Working on fixing layers that do not support 0/1D support
  - Alexander Panov:
    - Prepared PR Move Charuco/Calib tutorials and samples to main repo #3708 update/fix tutorials and samples
    - preparing updates for tutorial tests
    - Reviewed PRs: #3707, #25195
    - Works on findCirclesGrid can't detect some asymmetric grids #24971
  - Alexander Smorkalov:
    - Prepared 4.x->5.x merge #24893
    - Got rid of TIFF specific constants in imgcodecs module #25325
    - Updated CI scripts #167. Added CI pipeline for Windows ARM64 #155.
    - Reviewed, tested, helped to resolve issues ## 24881, 25087, 25209, 25301, 24877, 25273, 25299, 25293, 25271, 25280, 25291, 25255, 25092, 25282, 25177, 25287, 25259, 25221, 25322, 25252, contrib ## 3709, 3712, 3691
    - Works on #25034.
  - Alexander Lyulkov:
    - Fixed bugs in Tile and Reduce layers #25277
    - Added tests to ScatterND, Padding, Const, NaryEltwise, Arg, Blank, Concat layers

2024-03-27

Vadim:
- [OpenCV China] continue to work on transformer acceleration, removed unnecessary copies. Awaiting for GELU patch to be merged in.
- [OpenCV China] continue working on hfloat/bfloat universal intrinsics.
- [OpenCV China] OpenCV 5 cleanup goes on: samples, features2d+flann => features, darknet & caffe importers in dnn.
Vincent:
- javascript bindings; no reply on https://github.com/opencv/opencv/issues/25239 yet. Will OpenCV team look into it?
Phil:
Alexander S (opencv.ai):
- Back to work
- CVAT is going to move outside of github.com/opencv to their new home.
- ARM has published https://gitlab.arm.com/intrinsiccv/intrinsiccv. It supports ARM v8 and ARM v9 (including SVE2, bfloat etc.)
- Work on improving camera calibration is going on. Maxim Smolsky (from OpenCV community) has submitted a bunch of patches to improve Chessboard detector and other components of the calibration pipeline.
- Alexander Panov has improved circle detector.
- Aruco has been completely migrated from contrib to the main repository.
- Do we continue to work on calibration during GSoC 2024? If yes, need to decide fast, since April 2 is the deadline.

2024-03-20

Vadim:
- [OpenCV China] work on cleaning OpenCV goes on. PR on objdetect cleanup, Darknet & Caffe importers in DNN have been submitted.
- [OpenCV China] work on accelerating transformers goes on.
- Continue working on the new experimental DNN inference engine. The draft version will hopefully be ready in 2 weeks.
Vincent:
- Experimenting with bindings, JS bindings in particular. It looks like it does not work well currently. Will submit issue.
Phil:
- Plan to have OpenCV.ai/OpenCV.org at CVPR.
Gary:
- Interns are working on Gaussian splatting.
Shiqi:
- Can mentor some projects.
- Plan to attend CVPR this year.
Alexander S (opencv.ai):
- On vacation

2024-03-13

Vadim: *
Vincent:
- Would be nice to test OpenCV 5 alpha once it's out.
- no other items to discuss
Alexander S (opencv.ai):
- on vacation
Vladimir:
- OpenCV 5: Doxygen version used to generate opencv docs has been upgraded. Need to make some further changes.
- OpenCV 5: findContrours refactoring continues (in order to get rid of C API completely)
- PR was submitted with fixes for compare() function
- Reported bugs in T-head RISC-V compiler.
Dmitry:
- recent releases of OpenVINO broke OpenCV builds. Dmitry will provide patches.
- videocapture destructor (gstreamer backend) hangs from time to time. Will provide reproducer.

2024-03-06

Vadim:
- [OpenCV China] GELU activation function has been accelerated by 3x-6x. PR submitted.
- [OpenCV China] Wanli is working on replacing Darknet & Caffe models in our test data with ONNX, so that we don't have to support obsolete Caffe and Darknet in 5.0.
Vincent:
- ready to submit a patch (or a few) to conditionally disable exception handling in OpenCV (which can be useful for huge distributed apps that use OpenCV)
Satya:
- GSoC: need to adjust the ideas with OCV5.
- Maybe Gaussian splatting should be added to the list of ideas.
Alexander S (opencv.ai):
- Alexander Panov:
  - Merged PR Add strong transform to objdetect bench #12 review.
  - Improve contours approximations in ChessBoardDetector::generateQuads #25145
  - Prepared PR add to mcc detect and infer test #3647 added a full pipeline mcc tests
  - Prepared PR added getComputedColorCharts() #3645
  - Preparing Macbeth board improvements
  - Performance speedup to x4, the PR is being prepared
- Abduragim Shtanchaev:
  - ...
- Alexander Smorkalov:
  - Prepared regular 4.x->5.x merge (#25168).
  - Backport some C removal steps from 5.x to 4.x: ## 25160, 25149, 25125
  - Reviewed, tested, helped to resolve issues with ## 25162, 25161, 25146, 25148, 25123, 22997, 25141, 24411, 25120, 25122, 24724, 25096, 25075, contrib ## 3647, 3646, 3645
- Rostislav Vasilikhin:
  - Formulated a set of 1-month tasks for 3D module: #25150 #25151 25153 25154 #25155 #25156 #25157 #25158
  - There is also an old bug to fix #22896 and PRs to finish #22731 3D model benchmark:
  - Have chosen Scanned Objects by Google Research dataset. Implementation is in progress
- Alexander Lyulkov:
  - Fixed issue with int64 constants in ONNX parser (#25148)
  - Works on int32/int64 test coverage. Progress is tracked in the spreadsheet.

2024-02-28

Vadim:
- [OpenCV China] Feng Yuantao's patch on transformers acceleration has been merged: https://github.com/opencv/opencv/pull/25076. there is still a room for improvements.
- [OpenCV China] Wu Jia's patch to migrate ml from opencv to opencv_contrib is merged.
- Vadim is working on the new inference engine.
Vladimir:
- removing C API is in progress.
- reimplementing contours using C++ API: https://github.com/opencv/opencv/compare/4.x...mshabunin:cpp-contours.
Gary:
- GSoC: need to adjust the ideas with OCV5.
- Maybe Gaussian splatting should be added to the list of ideas.
Alexander S (opencv.ai):
- Macbeth chart detector: Alexander Panov is working on accelerating the detector. Planning to move it to the main repository in OpenCV 5
- Improving camera calibration that uses chessboard pattern.
- Asymmetric circle grid is not detected properly.

2024-02-21

Vadim:
- [OpenCV China] Feng Yuantao's patch on on-fly quantization removal in 5.0 is merged. We may (and likely will) restore in sometimes in OpenCV 5.x, but since we are now seriously refactoring DNN, such cleanup is quite useful.
- [OpenCV China] Wu Jia has submitted PR where ML module is moved from opencv to opencv_contrib, together with samples and docs.
- [OpenCV China] Zhong Wanli is working on adding v_exp() universal intrinsics.
- Vadim is working on the new inference engine.
Vincent:
- no updates
Alexander S (opencv.ai):
- Done:
  - Alexander Panov:
    - Merged PR Update QR formatting #9
    - Fixing issues in Migrate aruco tutorials
    - Prepared and merged PR Update non symmetrical patterns #11
    - Fixed and merged PR: Fix barcode detectAndDecode #25035
    - Prepared PR fix unnecessary copying in ArucoDetector #25045
    - Prepared PR fix setArucoParameters/getArucoParameters in QRCodeDetectorAruco #25046
    - Prepared and merged PR Update formatting and add seed to objdetect bench #10
  - Abduragim Shtanchaev:
    - Working on scatter layer 1d test. Finished layer tests
  - Alexander Smorkalov:
    - Prepared 4.x-5.x merge. Fixed conflicts.
    - Decomposed GSoC PR #24052 for mutli-camera calibration on several independent pieces ## 25061, 25022, 25021, 25023, 25028, 25030. Fixed review notes for the original code, added tests, documentation. The work is in progress.
    - Updated Gradle cache for Android builds with Wanli, made fixes in Android related CI (GHA #153, #152, Docker #33)
    - Reviewed, tested, helped to resolve issues with ## 23018, 24977, 24411, 25052, 25024, 25046, 24459, 24833, 25045, 25044, 25026, 25053, CI #132, benchmarks #11, 10, 9.
  - Rostislav Vasilikhin:
    - Triangle Rasterization PR: merged
    - Octree PR: remaking API, finishing
  - Alexander Lyulkov:
    - Working on finishing adding int support to dnn: fixed tests with float indices, fixed Mac arm support, fixed Python api bindings

2024-02-07

Vadim:
- [OpenCV China] Yuantao collected profile of ViT inference pipeline; looks like there are many opportunities to speed it up, like avoid repetitive memory allocation, optimizing GELU activation, optimizing SoftMax (maybe implementing FlashAttention), optimizing a few other layers besides matrix multiplication. Now Yuantao is also working on preparing CLBast-derived OpenCL MatMul and batch MatMul kernels for DNN (including the case when one of the matrix is constant).
- [OpenCV China] Jia has joined to active OpenCV 5 development. One of the activities will be OpenCV cleanup (moving obsolete parts to OpenCV contrib)
- Nearly finished PR with extending OpenCV arithmetics with FP16 support.
Dmitry M (Intel):
- working on updating G-API to fully support OpenVINO 2.0.
- immediate mode for G-API was made parallel.
Alexander S (opencv.ai):
- Done:
  - Alexander Panov:
    - Prepared and merged PR circle grid in OpenCV benchmark
    - Checking PR Do not dilate binary images on first iteration in findChessboardCorners
    - Fixing output for QR benchmark
    - Open issue findCirclesGrid can't detect some asymmetric grids #24971
    - Open issue findCirclesGrid can't detect rotated SYMMETRIC_GRID #24964
  - Abduragim Shtanchaev:
    - Worked on AndBool layer. Opened PR
    - Worked on ConvTranpose layer issue
  - Alexander Lyulkov:
    - Finished type inference proof-of-concept for cuDNN back-end added int support. Works on OpenCL code.
  - Alexander Smorkalov:
    - Prepared 4.x->5.x (#24958). The process stalled due to conflicts and issues in DNN.
    - Implemented offline mode for Android builds, added it to CI, built Gradle cache for offline mode in CI (GHA ## 145, 144, 141, opencv #24969).
    - Partially stabilized Android tests with real device. Looks like cable replacement is required.
    - Added OPENCV_TEST_REQUIRE_DATA option to CI to highlight missing test data issues (GHA #133). Added missing Open Model Zoo models (extra #1146), fix bug in DNN performance tests (#24973).
    - Reviewed, tested, helped to resolve issues with benchmarks #8 opencv ## 24621, 24411, 24779, 24459, 24949, 23985, 24197, contrib ## 3632, 3617.

2024-01-31

Satya:
- In the process of raising enough money to fund OpenCV 5 and beyond.
Vadim & OpenCV China:
- Looking at C++ libraries to run accelerated inference, especially transformer models.
- Started working on extending arithmetic functions to support F16, BF16, U32, U64 and S64.
- v_exp() has been implemented. Should be useful to accelerate various activation functions.
Vincent:
- Submitted getNewOptimalCameraMatrix(). The patch needs review.
Vladimir:
- Still fixing degradations in vectorized RISC-V code after updating compiler.
Alexander S (opencv.ai):
- Done:
  - Abduragim Shtanchaev:
    - Finished Yolo sample and tutorial text.
    - Working on ONNX conformance tests fixing for 1d cases.
  - Alexander Smorkalov:
    - Finished CI with Android Phone (CI #138).
    - Added warnings suppression list (#31), fixed several warnings found in OpenCV build for Android (#24945, #24942)
    - Reviewed, tested, helped to resolve issues with ## 24910, 24782, 24845, 24927, 24898, 24768, 24929, 24459, 24919, 24922, 24892
    - Prepared Android specific pages update for OpenCV.Org
  - Alexander Panov:
    - On sick leave
    - Works on circle grid in OpenCV benchmark better accuracy in pattern detection (0.16-0.2 pixels), problems with detecting the rotated non-square (when the number of rows is not equal to the number of columns) patterns.
  - Rostislav Vasilikhin:
    - Triangle rasterization PR from GSoC 2023 (#24459): ready for review again
      - Found an off-by-1 difference with ground truth on PI/4-sloped diagonal lines, failed to fix it, retained it with test thresholds
      - All parameters are taken in doubles now to prevent more precision shifts.
      - Tests are tightly fixed by a tree of thresholds for more corner cases
    - Octree PR #23985:
      - Reviewed, changes proposed
      - Rewriting according to proposed changes in progress
      - Other changes are postponed to next PRs
      - estimateAffine3D Umeyama bug:
      - Reproduction is in progress
  - Alexander Lyulkov:
    - Added Android and Kotlin tests #24910 #24946

2024-01-24

Vadim & OpenCV China:
- no major updates this week
Vincent:
- doing some work on libavif, going to finalize patches for OpenCV within next few weeks.
- ready to mentor GSoC project on animated image support (of course, if we are accepted, if there are good candidates etc.)
Alexander S (opencv.ai):
- started regular OpenCV 5 progress meetings
- Done:
  - Abduragim Shtanchaev:
    - Works on yolo documentation
    - Fixed Einsum layer 1d error
  - Alexander Smorkalov:
    - Prepared 4.x->5.x merge for all changes after 4.9.0 release. Branches are synchronized now.
    - Reviewed, tested, helped to resolve issues with ## 24910, 24459, 24782, 24873, 24900, 23018, 24906, 24899, 24898, 23736, 24904, 24878, contrib ##3615
    - Works on Core module API review. Filed #24908
  - Alexander Panov:
    - Works on circle grid in OpenCV benchmark
    - Works on PR Move Aruco tutorials and samples to main repo #23018
    - PR Fix generateImage for charuco chessboard #24873
    - Fixed and merged
  - Rostislav Vasilikhin:
    - Triangle rasterization PR from GSoC 2023 (#24459): ready for review again
      - API changed + GL compatible depth mode added
      - Decreased patch size by removing duplicated images and texture coordinates from OBJ file; still too big
      - In progress: making CI green
      - Will add K camera matrix in next PR
    - Point cloud compression PR. In progress:
      - #23985 should be merged first
    - Alexander Lyulkov:
      - Added kotlin code to Android AAR package #24884
      - Works on Android JUnit tests with Gradle and real device #24910

2024-01-17

Vadim & OpenCV China:
- [OpenCV China] Yuantao has tested CLblast on different machines. Normally, CLBlast's gemm is 1.15x-5x faster than OpenCV's OpenCL gemm (and faster than the new fast CPU gemm).
- [OpenCV China] Optimization of transformers goes on. Yuantao has profiled VIT inference. Apparently, GELU takes >10%; there is also surprisingly expensive Constant node (2-3%) and Transpose node (that can be fused with preceding or following MatMu/GEMM).
- [OpenCV China] Wanli is working on replacing Darknet models with ONNX. Yolo models have been converted and seemingly work well, but post-processing in the tests needs to be updated to make the tests pass.
Gary:
- GSoC 2024: application period starts next week. Need to prepare the list of ideas.
- Q: new camera calibration - what is it about? does it support Charuco as calibration rig? A by Alexander S: in theory, Charuco is supported, but there are some issues with it, it still requires some work.
Vincent:
- no big updates; opened issue related to animation support (https://github.com/opencv/opencv/issues/24855).
Vladimir D:
- Finally finished with 1D signal resampling func in OpenCV contrib.
- Additionally, working on integration of T-Head changes to support their RISC-V toolchain 2.8.1
Dmitry M (Intel):
- asked question about customizing CI jobs (e.g. to test fresh OpenVINO)
Alexander S (opencv.ai):
- Abduragim Shtanchaev: on vacation
- Done:
  - Alexander Smorkalov:
    - Prepared 4.x -> 5.x merge for commits up to 4.9.0
    - Debugged and fixed out-of-memory access in Yolo-NAS test (#24809, 24861)
    - Prepared OpenCV 5.x core update plan for new types support.
    - Reviewed, tested, helped to resolve issues with ## 24869, 24868, 24860, 24773, 24813, 24873, 24874, 24768, 24459, 24567, 24864, 24856, 24852, 24849, 24842, 24812, 24830 contrib ##3618, 3615, 3616
  - Alexander Panov:
    - Prepared PR Fix generateImage for charuco chessboard #24873
    - Improve accuracy of generation chessboard, added test
    - Review and commented #24558
    - Works on moving Aruco tutorials and samples to main repo #23018
  - Rostislav Vasilikhin:
    - Finalized rasterization PR from GSoC 2023 (#24459)
    - Works on Point Could Compression PR (#24197)
  - Alexander Lyulkov:
    - Added screen rotation support to Android JavaCameraView, JavaCamera2View and NativeCameraView. Added screen rotation support to tutorial1 and face-detection Android samples.
    - Added javadoc compilation in AAR creating script
    - Works on Kotlin dependency in SDK and published AAR packages

2024-01-10

Vadim & OpenCV China:
- Several PRs with DNN improvements and bug fixes have been submitted.
- Found very interesting CLblast project - will possibly use it to accelerate OpenCL backend of OpenCV DNN.
- There is "sort-of-successful" attempt to run CLIP model using OpenCV. 1/2 models (the visual part) is run using OpenCV, the other model (textual part) is now running using ONNXruntime. Can be a nice project to enable the whole thing in OpenCV.
- Looking for ways to provide better out-of-the-box performance of OpenCV Python @ Intel.
- GSoC 2024 is approaching, time to start preparations.
Gary:
- working on defining the next directions for OpenCV.
Vincent:
- 4.9.0 has been tested internally.
- working on animation support in webp codec in OpenCV.
Alexander (opencv.ai):
- Done:
  - 4.9.0 is out just before new year. The changelog is here, as usual: https://github.com/opencv/opencv/wiki/ChangeLog.
  - Abduragim Shtanchaev:
    - Finished 1d Vector operation support in Einsum. PR #24830
    - Added FastBatchedGemm for Einsum. PR #24812
    - Added yolo-nas test PR #24309. Works on memory corruption issue in parser.
    - Solved issue with concat layer on CUDA #24786
  - Alexander Smorkalov:
    - Reviewed, tested, helped to resolve issues with ## 24829, 24765, 24799, 24828, 24823, 24819, 24820, 24783, 24815, 24801, 24807, 24814
    - Prepared fixes for Android ## 24788, 24843, 24842
- Alexander Panov:
  - Fixing issue ChArUco generation produces incorrect marker locations in the Chessboard #24806
  - Reviewed and tested PR Fix QR code black/white balance computation #24758, PR * Make aruco detector deterministic #24829
  - Fixed and merged PR Update android OpenCL sample #24715
- Rostislav Vasilikhin:
  - Triangle rasterization PR #24459:
    - Test data generator finished
    - Accuracy & perf tests refactored and almost done
    - In progress: source refactoring, args checking and corner cases
    - In progress: changing API to more convenient