Can CascadeClassifier.detectMultiScale be run in several threads at once?

[Benjamin Xiao]

Hello again, fellow JavaCV users!

So I am using CascadeClassifier to do object detection in a video feed. Currently I am doing upper body, full body, and face detection in 3 separate CascadeClassifier instances. I am trying to run each detection in a separate CPU thread. However, that ends up being slower than if I run them sequentially. It almost seems like there's some lock behind calls to CascadeClassifier.detectMultiScale that's only allowing one call to run at a time.

Here is how I am testing this:

// Create classifier jobs

var classifierJobs = getClassifiers().stream()

        .map((classifier) -> {

            Callable<List<ClassifierResult>> classifierJob = () -> {

                var detectedObjs = new RectVector();

                classifier.detectMultiScale(grayMat,

                    detectedObjs,

                    scaleFactor,

                    minNeighbors,

                    CASCADE_SCALE_IMAGE,

                    new Size(minSize, minSize),

                    new Size()

                );

                return detectedObjs;

            };

            return classifierJob;

        })

        .collect(Collectors.toList());

// Execute jobs with ExecutorService instance

ExecutorService classifierPool = Executors.newCachedThreadPool();

classifierPool.invokeAll(classifierJobs)

As you can see, I am creating a Callable for each CascadeClassifier instance. Inside, there's a call to detectMultiScale. I feed all of these Callables to a cached thread pool via invokeAll.

This code runs but it is slow. If I instead define classifierPool as newSingleThreadExecutor() then the code runs much faster, albeit in a sequential fashion on a single thread.

So what's the deal here? Am I running into threading overhead? I thought using a thread pool would decrease some of the overhead particularly in terms of thread creation and destruction.

Or is there some kind of mutex lock behind the scenes that I am not aware of that's causing my multithreaded code to slow down?

Any help appreciated. Thanks in advance!

Ben

[Samuel Audet]

OpenCV already uses all threads for each call. If you'd like to disable multithreading in OpenCV, you can try to call setNumThreads(1).

Samuel

[Benjamin Xiao]

Thanks Samuel! Did not know that detectMultiScale was already multithreaded. That probably explains why I am running into the issues I am experiencing. I guess I'll implement a job queue and feed these classifier instances one by one.

Ben