Does Opencv3.4 support Mali-T860 GPU acceloration?

asked 2018-07-16 22:55:04 -0600

bryan2lee gravatar image

updated 2018-07-17 00:39:55 -0600

My Development Platform:

  • Android7.0 (abi-armeabi-v7a)with android studio
  • The OpenCV version is 3.4
  • Feature: UVC CAMERA + OPENCV(For Face Detection)
  • I have already successfully run uvc withopencv on my board, but the performance is not as well as expetected
  • All this set of code runs on CPU

int flags = cv::CASCADE_FIND_BIGGEST_OBJECT|cv::CASCADE_DO_ROUGH_SEARCH; 

cv::Size minFeatureSize(30, 30);

float searchScaleFactor = 1.1f; 

int minNeighbors = 2; 

faceDetector.detectMultiScale(equalizedImg, faces, searchScaleFactor, minNeighbors, flags, minFeatureSize);

  • So I wonder if it's possible to run faceDetector.detectMultiScale on GPU(Mali-T860)
  • I checked the source file ocl.hpp, It seems that OpenCV only supports 3 types of GPU for now

enum
    {
UNKNOWN_VENDOR=0,
        VENDOR_AMD=1,
        VENDOR_INTEL=2,
        VENDOR_NVIDIA=3
    };
    int vendorID() const;

So here comes my Question: Is it possible to run face detection on Mali-T860(it supports opencl 1.2) If the answer is yes, then how to do it? There is no any answers on the opencv.org

edit retag flag offensive close merge delete

Comments

i cannot answer your question, but please also check out the dnn based object detection . there is a pretrained model for faces, too, and it'S much faster than the cascade based one.

berak gravatar imageberak ( 2018-07-17 00:50:25 -0600 )edit

Thanks anyway So do you mean that dnn can also be used for face-detecting?

bryan2lee gravatar imagebryan2lee ( 2018-07-17 00:59:12 -0600 )edit

yes, exactly. it's just a different model

berak gravatar imageberak ( 2018-07-17 01:30:25 -0600 )edit

Does this model run on CPU or GPU?

bryan2lee gravatar imagebryan2lee ( 2018-07-17 01:32:08 -0600 )edit

there is an opencl backend, but i have no idea about the state of it on android

berak gravatar imageberak ( 2018-07-17 01:44:03 -0600 )edit
1

its pretty fast even on cpu (that caffe face detector model) - around 2,5 ms - give it a try and measure by yourself.

holger gravatar imageholger ( 2018-07-17 23:59:36 -0600 )edit