search

fastfastball / arm_neon_for_dlib_simd

add arm neon support to dlib. now, merged into https://github.com/fastfastball/dlib_for_arm
22 stars 4 forks source link

Something about how to enable neon option #3

Open lazylazypig opened 8 years ago

lazylazypig commented 8 years ago

Hi, it's a amazing work! i would like to know how to enable neon when i build dlib using cmake. my platform is nvidia jetson tk1, with ARM Cortex-A15 CPU.

fastfastball commented 7 years ago

you may check this link https://github.com/fastfastball/dlib_for_arm/issues/1

gangm commented 7 years ago

@lazylazypig , hello, we are using TK1 too, but the performance is a little slow, can you tell us how long do you get in TK1 to detect one picture? we used 800*600 picture, and it took about 1.8s to detect one picture, i think it is too slow, any other detail to be noticed?

our code: frontal_face_detector detector = get_frontal_face_detector(); load_image(img, argv[i]); for(int count=0; count < 100; count++) { double t1 = cv::getTickCount(); cout << "start to detect..." << endl; std::vector dets = detector(img); cout << "Number of faces detected: " << dets.size() << endl; double t2 = cv::getTickCount(); std::cout << "Read time: " << (t2 - t1) * 1000 / cv::getTickFrequency() << " ms." << std::endl; }

the result: processing image face.jpg start to detect... Number of faces detected: 7 Read time: 1808.41 ms. start to detect... Number of faces detected: 7 Read time: 1815.63 ms.

lazylazypig commented 7 years ago

@gangm hi, sorry for my late reply. it's been a long while since i tested dlib on tk1, so i cannot remember the exact time cost for detecting one picture. but, i am sure that the performance is not very different from yours. because of the limit of performance, finally i have to give up using dlib on tk1.