Question about the speed of face detector between 3.3 and 3.4

asked 2018-02-27 02:11:57 -0600

loogles gravatar image

updated 2018-02-27 02:41:43 -0600

berak gravatar image

Hi, guys:) I used the resnet_ssd_face.cpp provided in version 3.4 and successful to detect faces in video stream. However, I found that when I use the library in 3.3, the speed is much faster then the in 3.4 (about 3~4 times). I want to know how this happened? All I did is just change the project include path / library path / link. Thanks.

BTW: I download the winpack version of 3.3.0 and 3.4.0 from https://opencv.org/releases.html

edit retag flag offensive close merge delete

Comments

1

My guess, they introduced the Halide accelerator backend in 3.4 and from the inference timing page it seems to be slower than the pure CPU version. @dkurt can you elaborate on this maybe?

StevenPuttemans gravatar imageStevenPuttemans ( 2018-02-27 03:19:05 -0600 )edit
1

@loogles, please add a build configurations using cv::getBuildInformation and some absolute efficiency measurements (in milliseconds or a number of frames per second) including hardware description.

dkurt gravatar imagedkurt ( 2018-02-27 06:10:11 -0600 )edit
add a comment