On FPGA (but not CPU), face detection plus add-on models throws resize error.

[SGrif4]

Hardware/Software Used

Intel® Distribution of OpenVINO™ Toolkit 2018 R5

Intel® Vision Accelerator Design with Intel® Arria® 10 FPGA (IEI's Mustang-F100-A10)

On FPGA (but not CPU), face detection plus add-on models throws resize error.

following this: https://github.com/fritzboyle/openvino-with-fpga-hello-world-face-detection

Takeaways: (Looking at FPS of one video at a time using these models).

FPGA is not able to handle face detection plus add-ons (like age gender or head pose) on any video that was NOT the sample demo video as it runs into a resize error (highlighted below in yellow). However the CPU can run the face detection model with add-ons on any video (including the demo video).

Comparing the CPU vs FPGA running just the face detection model, the FPGA was always slightly slower than the CPU.

Questions/Concerns:

·     Why won’t face detection plus add-on models work on non-demo videos on the FPGA? (As in, will work on demo video that comes with tutorial will support add-on face detection, but add-on face detection won't work on my own video file).

One video – demo video

W/out Specifying FPGA (aka CPU) –

[ INFO ] Total main-loop time:80316.55 ms

[ INFO ] Total # frames:1719

[ INFO ] Average time per frame:46.72 ms (21.40 fps)

HETERO mode with FPGA –

[ INFO ] Total main-loop time:55428.89 ms

[ INFO ] Total # frames:1719

[ INFO ] Average time per frame:32.24 ms (31.01 fps)

One video – non demo video

W/out Specifying FPGA (aka CPU) –

[ INFO ] Total main-loop time:88910.86 ms

[ INFO ] Total # frames:2029

[ INFO ] Average time per frame:43.82 ms (22.82 fps)

HETERO mode with FPGA –

[ ERROR ] OpenCV(4.0.1-openvino) /home/jenkins/workspace/OpenCV/OpenVINO/build/opencv/modules/imgproc/src/resize.cpp:3784: error: (-215:Assertion failed) !ssize.empty() in function 'resize'

[JohnT_Intel]

Hi,

Comparing the CPU vs FPGA running just the face detection model, the FPGA was always slightly slower than the CPU.

What do you mean by this? The result that you shared, I am seeing that with CPU only is running 21fps while with FPGA then it is able to run up to 32fps. This is showing with FPGA, it is able to run faster.

Questions/Concerns:

·     Why won’t face detection plus add-on models work on non-demo videos on the FPGA? (As in, will work on demo video that comes with tutorial will support add-on face detection, but add-on face detection won't work on my own video file).

May I know if you are able to share with me the video? It should be able to work as the image resize should be performed in CPU.

[SGrif4]

Comparing the CPU vs FPGA running just the face detection model, the FPGA was always slightly slower than the CPU.

What do you mean by this? The result that you shared, I am seeing that with CPU only is running 21fps while with FPGA then it is able to run up to 32fps. This is showing with FPGA, it is able to run faster.

---> Right, my error typo in the note. Thanks for pointing this out.

Questions/Concerns:

·     Why won’t face detection plus add-on models work on non-demo videos on the FPGA? (As in, will work on demo video that comes with tutorial will support add-on face detection, but add-on face detection won't work on my own video file).

May I know if you are able to share with me the video? It should be able to work as the image resize should be performed in CPU.

--> Please find it attached below!

[JohnT_Intel]

Hi,

May I know if it is possible to upgrade your OpenVINO to 2019R1? I have tested your video and it is able to run on my setup without any issue.