Open CV will not infere more than 2 models when cv::utils::logging::setLogLevel(cv::utils::logging::LOG_LEVEL_VERBOSE); is set

[ukoehler]

System Information

OpenCV 4.7.0 and 4.8.0 compiled from source
Operating system: various Linux versions
Compiler: gcc5, gcc7 and gcc11

Detailed description

Took me a whole week to investigate #23982 and found a very bizzare behaviour that nobody will easily expect.

When infering DNNs everything is ok, as long as I infere the same model several times (I tried up to 3). However, the problem start when infering different models. With standard logging, still, everything works fine.

However, when cv::utils::logging::setLogLevel(cv::utils::logging::LOG_LEVEL_VERBOSE); is used (I use that to ensure my build of OpenCV is actually multi-threading on various platforms), the third inference stops working reproducibly . The provided code as 4 net and commenting inferences shows that the concrete model does not seem to matter.

Steps to reproduce

The problem is reproducible and enforceable.

The code used:

#include "opencv2/opencv.hpp"
#include "opencv2/core/utils/logger.hpp"

        cv::utils::logging::setLogLevel(cv::utils::logging::LOG_LEVEL_VERBOSE);
        std::string alexnetModelConfiguration = R"(/build/uk/source/ovtest/deploy_alexnet_places365.prototxt)";
        std::string alexnetModelWeights = R"(/build/uk/source/ovtest/alexnet_places365.caffemodel)";

        // Load the network
        cv::dnn::Net alexnetNet = cv::dnn::readNetFromCaffe(alexnetModelConfiguration, alexnetModelWeights);
        alexnetNet.setPreferableBackend(cv::dnn::DNN_BACKEND_OPENCV);
        alexnetNet.setPreferableTarget(cv::dnn::DNN_TARGET_CPU);

        std::string googlenetModelConfiguration = R"(/build/uk/source/ovtest/deploy_googlenet_places365.prototxt)";
        std::string googlenetModelWeights = R"(/build/uk/source/ovtest/googlenet_places365.caffemodel)";

        // Load the network
        cv::dnn::Net googlenetNet = cv::dnn::readNetFromCaffe(googlenetModelConfiguration, googlenetModelWeights);
        googlenetNet.setPreferableBackend(cv::dnn::DNN_BACKEND_OPENCV);
        googlenetNet.setPreferableTarget(cv::dnn::DNN_TARGET_CPU);

        std::string resnet152ModelConfiguration = R"(/build/uk/source/ovtest/deploy_resnet152_places365.prototxt)";
        std::string resnet152ModelWeights = R"(/build/uk/source/ovtest/resnet152_places365.caffemodel)";

        // Load the network
        cv::dnn::Net resnet152Net = cv::dnn::readNetFromCaffe(resnet152ModelConfiguration, resnet152ModelWeights);
        resnet152Net.setPreferableBackend(cv::dnn::DNN_BACKEND_OPENCV);
        resnet152Net.setPreferableTarget(cv::dnn::DNN_TARGET_CPU);

        std::string vgg16ModelConfiguration = R"(/build/uk/source/ovtest/deploy_vgg16_places365.prototxt)";
        std::string vgg16ModelWeights = R"(/build/uk/source/ovtest/vgg16_places365.caffemodel)";

        // Load the network
        cv::dnn::Net vgg16Net = cv::dnn::readNetFromCaffe(vgg16ModelConfiguration, vgg16ModelWeights);
        vgg16Net.setPreferableBackend(cv::dnn::DNN_BACKEND_OPENCV);
        vgg16Net.setPreferableTarget(cv::dnn::DNN_TARGET_CPU);

        std::string inputFile = R"(/build/uk/source/ovtest/Beach.png)";

        cv::Mat image = cv::imread(inputFile, cv::IMREAD_UNCHANGED);

        cv::Mat blob227 =
            cv::dnn::blobFromImage(image, 1.0, cv::Size(227, 227), cv::Scalar(103.94, 116.78, 123.68), true, false);
        cv::Mat blob224 =
            cv::dnn::blobFromImage(image, 1.0, cv::Size(224, 224), cv::Scalar(103.94, 116.78, 123.68), true, false);

        std::vector<cv::Mat> ret;
        alexnetNet.setInput(blob227);
        alexnetNet.forward(ret);
        std::cout << ret.size() << "  " << ret[0].rows << "  " << ret[0].cols << std::endl;

        googlenetNet.setInput(blob227);
        googlenetNet.forward(ret);
        std::cout << ret.size() << "  " << ret[0].rows << "  " << ret[0].cols << std::endl;

        resnet152Net.setInput(blob224);
        resnet152Net.forward(ret);
        std::cout << ret.size() << "  " << ret[0].rows << "  " << ret[0].cols << std::endl;

        vgg16Net.setInput(blob224);
        vgg16Net.forward(ret);
        std::cout << ret.size() << "  " << ret[0].rows << "  " << ret[0].cols << std::endl;

Error message:

OpenCV(4.8.0) Error: Requested object was not found (Required argument "operation" not found into dictionary) in get, file /build/uk/source/4.8.0_trial/source/lib/modules/dnn/include/opencv2/dnn/dnn.inl.hpp, line 350
[DEBUG:0@3.570] global system.cpp:2881 restoreFPDenormalsState core: restore FP mxcsr flags = 0x00001fa0
ERROR: caught an exception

I used the following files for the nets:

• https://drive.google.com/file/d/1Q6rOuY3XW0EB1QIFb2xpcsXLhjir22rF/view?usp=drive_link
• https://drive.google.com/file/d/1-f_S2o76-NjH0_Ke32JqgetX7z-xrPKr/view?usp=drive_link
• https://drive.google.com/file/d/1ZkXRvD8rEhxXMxLpfPvukd-OPxUsdCwK/view?usp=sharing
• https://drive.google.com/file/d/1fi7ETFWeZEif-nh6za0sxLXvudtfR5Ib/view?usp=sharing
• https://drive.google.com/file/d/1yZwED-J3--lxwMgUGJfa5-7ONtLqTyP4/view?usp=sharing
• https://drive.google.com/file/d/1ACj0_BKF4dENsV1W3Jpb2bOXZhYp1d74/view?usp=sharing
• https://drive.google.com/file/d/12DRxFjHI8CD1Mmd7fyQrGUW8GnV_kw58/view?usp=sharing
• https://drive.google.com/file/d/1A7QhfnTn9RJK0--EOS0SMGroZR4gXJfG/view?usp=sharing

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 @ukoehler, update: test with MacOS 12.1, intel i9 and it works well, compier is XCode. Will test it on my Ubuntu later.

[ukoehler]

Did you test release or debug?

[zihaomu]

Did you test release or debug?

release

[ukoehler]

Your are not testing the problem: According to the documentation cv::utils::logging::setLogLevel(cv::utils::logging::LOG_LEVEL_VERBOSE); cannot be used in release mode. It is not documented what will happen if you use it anyway.

You have to build a debug version to see the effect.

[zihaomu]

@ukoehler. Thanks for your feedback. I'm working on it.