High CPU usage

[Q3Varnam]

I am comparing performance of OpenIMAJ against OpenCV(using Java).

I am running a simple face detector using haarcascade_frontalface_default.xml as given in OpenIMAJ tutorial.

I am running it on a dual core AMD A6 laptop with 6GB RAM

When I run OpenIMAJ version the CPU load is 120% when its detecting faces and 110% when just displaying a video without any faces.

When I run the OpenCV Java version of the face detector, CPU load is at 60% when detecting faces and 40% when the video without faces is displayed .

All measured using top command.

On OpenIMAJ, I have set the frame rate to 4.

On OpenCV I am using a thread which sleeps for 250ms before it captures a new image from the camera.

The scaling is set to 1.1 on both OpenCV as well as OpenIMAJ

On the OpenCV version I am additionally calculating the LBPH histograms before displaying the image.

I tried forcing OpenIMAJ program to use the GPU by passing -Dprism.forceGPU=true -Dprism.verbose=true but that doesn't make any difference.

I am running this on Ubuntu 14.04 with Java 8.

Can you confirm whether this is the normal CPU usage for OpenIMAJ or whether I am missing something which is unnecessarily loading the CPU?

[Jonathon Hare]

This is likely to be something to do with the way the native drivers for your camera handle requests for specific frame rates. Instead of specifying the framerate in your VideoCapture constructor, you could just do the same as you're doing with OCV and sleep the thread:

final VideoCapture vc = new VideoCapture(640, 480);
final HaarCascadeDetector hcd = HaarCascadeDetector.BuiltInCascade.frontalface_default.load();

//Either do it this way:
// for (final MBFImage img : vc) {
// for (final DetectedFace f : hcd.detectFaces(img.flatten())) {
    // img.drawShape(f.getBounds(), RGBColour.RED);
    // }
    // DisplayUtilities.displayName(img, "foo");
    //Thread.sleep(250);
// }

//or this way:
VideoDisplay.createVideoDisplay(vc).addVideoListener(new VideoDisplayAdapter<MBFImage>() {
@Override
public void beforeUpdate(MBFImage frame) {
    for (final DetectedFace f : hcd.detectFaces(frame.flatten())) {
    frame.drawShape(f.getBounds(), RGBColour.RED);
    }
    try { Thread.sleep(250); } catch (final InterruptedException e) {}
}
});

In my tests I'm seeing ~55% CPU usage with either of these approaches (and getting similar results to you (~120% usage) if I ask the native camera drivers to provide me with a 4 fps stream). JVisualVM indicates that the bit taking the most CPU time in the case where the frame rate is set to 4 fps is in the native call to grabber.getImage() inside VideoCapture.getNextFrame().

Note that if you want to improve the performance of the face detector, setting the minimum detection size to be larger is a good place to start.

[Q3Varnam]

With your suggestion things have improved far better, your code outperforms OpenCV while using most of the cascade files,

Cascade file used	OpenIMAJ CPU Usage	OpenCV(Java) CPU
frontalface_default	20	50
frontalface_alt	12	40
frontalface_alt2	12	40
frontalface_alt_tree	55	40

I am not sure why the scroll bar is not appearing for the table

Unformatted values are below if you can't read the table

frontalface_default 20 50
frontalface_alt 12 40
frontalface_alt2 12 40
frontalface_alt_tree 55 40