Axis Customer Forum

It is currently Sun Sep 15, 2019 5:02 pm

All times are UTC




Post new topic Reply to topic  [ 1 post ] 
Author Message
PostPosted: Mon Apr 08, 2019 1:57 pm 

Joined: Fri Sep 21, 2007 2:03 pm
Posts: 7
There is a lot of talking about deep learning in our industry but it is sometimes difficult to demonstrate deep learning in action with Axis cameras. IPVM published their AI Video Tester which is a very interesting initiative. The downside is that you need to upload videos which might be difficult when you have sensitive material.

Hence I have published two simple applications:
• Face recognition based on Dlib
• Object identification based on Darknet’s YOLO3
You will find those applications here: http://ftp.axis.com/pub_soft/SolutionsLab/DeepLearning/
They are provided “as is” for free without any promises of support.

The applications are implemented in Python and that makes installation a little bit tricky although I have tried to make it as simple as possible you might experience some obstacles. And, the good old command line is back to control the behavior of the software. On the positive side, it is quite easy to modify the code in case something is missing. Some additional notes:

Face recognition
Dlib is an amazing library and gets you decent results for both face detection and face recognition. You will need some good reference images to get it to work correctly. I usually prefer pictures taken by the camera that is used for recognition. Use frontal head shots. Face recognition requires a lot of processing power and you might experience additional latency when you have lots of faces in the video or faces in the video for longer time. You might want to experiment with resolution and fps depending on your computer. The detection threshold is set by default to 0.3 which gives good results. Lower the value somehow if you get too many false recognitions.

Object Identification
YOLO3 is one of the best object identifications algorithms that have been published. It detects 80 types of objects – people, vehicles, animals and household items. It requires lots of processing power and it helps a lot to use an nVidia GPU. Hence there are two versions for the installation. YOLO3 analyzes frames of a size of 416x416 pixels so there is no need to pull video streams of high resolution 640x480 or 800x450 work very well.

Here are some performance measurements taking into account only the inference part of the algorithm:

CPU 768ms (i5-8250 @ 1.6GHz)
GPU 23ms (i7-8700 @ 3.2GHz + nVidia GeForce GTX 1080 Ti)
TX2 268ms (nVidia Jetson YX2 based on Tegra X2)
NCS2 830ms (RPi + Intel NCS2)

If your computer has not enough performance, you now have an excuse to ask for a hardware upgrade.

Please let me know if you have any questions or comments.


Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 1 post ] 

All times are UTC


You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum

Search for:
cron
Powered by phpBB® Forum Software © phpBB Group