Fix missing log_gamma in nfa() #23052
Conversation
DNN: Make Unsqueeze layer support negative axes
Parsing quantized nodes does not rely on names
cosmetic changes
DNN: supports Scatter and ScatterND from ONNX
Setting CAP_PROP_AUTO_EXPOSURE on VideoCapture with backend DSHOW does not change anything. Now with this implementation the property can be used with value 1 for availability.
Trigger on "category: dnn (onnx)" label
Address CUDA-related errors
optimize winograd futher more
…daptiveThreshold QR code, reduce extra adaptiveThreshold()
* docs: remove duplicated step in Build .js doc * docs: add missing emcmake prefix to command
* Fixes for Halide * Enable some Halide tests
DNN-TF: let StridedSlice layer support const input
DNN: bug fixed in Winograd
android: add android:exported="true"
android: update build scripts
Do not fail tests in Yolo v7 model was not found
|
Hi, I saw that the test failed always when testing with the rotatedRect: opencv/modules/imgproc/test/test_lsd.cpp Line 156 in 6a7d54f I did some testing to try to understand why it was failing, but then I tested the same test with the original LSD method it was also failing the test. The reason why the incorrect method is working is that doing the refinement with LSD_REFINE_ADV returns almost the same result as using LSD_REFINE_STD because rect_improve is basically almost doing nothing. The issue is that since the log_gamma is missing the value of opencv/modules/imgproc/src/lsd.cpp Line 1077 in 6a7d54f Here is the calculations with the current code for one of the lines of the rotated rect test: And this are the same calculations with the corrected code: As a result, with the current code, it in many cases (if not always), the if statement below becomes true, and therefore the value of nfa in the current code is either opencv/modules/imgproc/src/lsd.cpp Lines 1079 to 1083 in 6a7d54f
Edit:
|
* fix openmp include and link issue on macos * turn off have_openmp if OpenMP_CXX_INCLUDE_DIRS is empty * test commit * use condition HAVE_OPENMP and OpenMP_CXX_LIBRARIES for linking * remove trailing whitespace * remove notes * update conditions * use OpenMP_CXX_LIBRARIES for linking
Add `estimateSE2(...)`, `estimateSE3(...)`, `estimateSIM2(...)`, `estimateSIM3(...)` for estimating an geometric transformation with rotation and translation (with scaling for SIM) using USAC: as alternative for `estimateAffinePartial2D` and `estimateAffine3D`. Modified test module. Remove unused variables. Remove initializer of unused variable. Add interfaces to accept UsacParams() and corresponding test codes. Revise test code. PartialNd removed Umeyama rewritten for code quality & speed comments & minors rise number of points fix, and +30% faster! only one number should be that big remove USAC code, leave fix only big number
…ense USAC fix: GraphCut fails to allocate big dense matrices
Adding HEVC/H265 FourCC support to MSMF video writer * Adding HEVC/H265 fourcc to MSMF video writer Adding HEVC/H265 fourcc to MSMF video writer. I have verified it with my own video input stream, and it works well on my workstation. * Update video io testing * Adding macro fence to get rid of compiler error H265/HEVC encoder is only available in Windows or later. https://learn.microsoft.com/en-us/windows/win32/medfound/h-265---hevc-video-encoder * Update test_video_io.cpp
Added regression parameterized test for Structure Append mode final_qr_code clear outside generateQR() method
dnn: add layer normalization for vision transformers * add layer norm onnx parser, impl and tests * add onnx graph simplifier for layer norm expanded * handle the case when constants are of type Initializer * add test case for layer norm expanded with initializers * use CV_Assert & CV_CheckType in place of CV_Assert_N; use forward_fallback for OCL_FP16 * use const ref / ref in parameters of invoker::run; extract inner const if from nested loop; use size_t in place of ull * template hasBias * remove trailing whitespace * use pointer parameter with null check; move normSize division & mean_square division outside of loop; use std::max to ensure positive value before std::sqrt * refactor implementation, optimize parallel_for * disable layer norm expanded * remove the removal of layer norm optional outputs
…_23105_encodeStructuredAppend_problem Fix encodeStructuredAppend() resulting in only one QR code problem, and false output data fix.
|
I have found the other issue (there might be more). The rect_nfa function was not checking the pixels inside the rectangle correctly. With the last changes to the rect_nfa function, now it checks the same pixels as the original paper. In the image below I fill a rotated rectangle using the original paper method (blue), the old OpenCV method (green), and the fixed code (red). As you can see, the old calculation had some logic issues. I have tested the fix with the rotatedRect test example, and now it detects all the lines properly (compared to before when sometimes some of the lines were missing). |
|
Closing PR and opening a new one due to a mistake merging opencv:4.x |
Comparing the nfa function with the function in the binomial_nfa repository, the first log_gamma call is missing.
Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
Patch to opencv_extra has the same branch name.