Segmentation fault for cv::dnn::readNetFromONNX.forward()

[UNeedCryDear]

System Information

OpenCV version: 4.7.0 from https://github.com/opencv/opencv/archive/refs/tags/4.7.0.zip
Operating System / Platform: Win10 22H2
Compiler & compiler version:VS2019 -MSVC14.2

Detailed description

ONNX export from yolov5-seg.pt:
https://drive.google.com/file/d/1tV2moQxNfLzNf6yXm5Zev5CVj2o9zuaz/view?usp=share_link

Some wrong happened in runWinograd63()

opencv/modules/dnn/src/layers/fast_convolution/fast_convolution.cpp

Line 420 in 9208dcb

if (runWinograd63(input, fusedAddMat, output, conv, ntasks, minval, maxval, activ, ifMinMaxAct))

Wwhen running here finally, an error is reported: the vector is out of range.

opencv/modules/dnn/src/layers/fast_convolution/winograd_3x3s1_f63.cpp

Line 1047 in 9208dcb

parallel_for_(Range(0, ntasks), [&](const Range& r0) {

@lioneldaniel told me that it was caused by this PR #22593, so how to solve this problem？

UNeedCryDear/yolov5-seg-opencv-onnxruntime-cpp#6 (comment)

Steps to reproduce

test code:

#include <iostream>
#include <opencv2/opencv.hpp>
using namespace cv;
using namespace std;
string model_path = "./yolov5s-seg.onnx";
Net net=readNetFromONNX(model_path );
net.setPreferableBackend(cv::dnn::DNN_BACKEND_DEFAULT);
net.setPreferableTarget(cv::dnn::DNN_TARGET_CPU);
Mat test_img=Mat::zeros(Size(640,640),CV_8UC3);
Mat blob;
blobFromImage(test_img, blob, 1 / 255.0, cv::Size(640, 640), cv::Scalar(0, 0, 0), true, false);
net.setInput(blob);
vector<string> output_layer_names{ "output0","output1" };
net.forward(net_output_img, output_layer_names); 
...

The above code can run correctly in versions 4.5.0~4.6.0, but an error occurred in 4.7.0.

Issue submission checklist

• I report the issue, it's not a question
• I checked the problem with documentation, FAQ, open issues, forum.opencv.org, Stack Overflow, etc and have not found any solution
• I updated to the latest OpenCV version and the issue is still there
• There is reproducer code and related data files (videos, images, onnx, etc)

[zihaomu]

Hi @UNeedCryDear, I just check your model with 4.7 release. And it works fine on my ARM M1 chip and Ubuntu of X86 platform (with AVX2).

I will try to reproduce the issue in Win10 later.

For short you can try to close the Winograd optimization in your machine by the following code:

net.enableWinograd(false);
net.forward(net_output_img, output_layer_names); 

Mode details please refer the doc link: https://docs.opencv.org/4.x/db/d30/classcv_1_1dnn_1_1Net.html#a14a87a7604c03ef4ff366672ee9bfcf2

[UNeedCryDear]

It can run correctly after adding net.enableWinograd(false);

[zihaomu]

Hi @UNeedCryDear, the patch has been merged. You can try to rebuild OpenCV from scratch to get the speed up of Winograd in your machine.

[lioneldaniel]

@zihaomu I confim that the following code works now, thanks!

opencv/modules/dnn/src/layers/fast_convolution/fast_convolution.cpp

Line 141 in 65c2d6a

else if(conv->conv_type == _FX_CONV_TYPE_WINOGRAD3X3) // winograd