This is a continuation of results of my performance testing. I realize that I'm collecting a huge amount of data and it doesn't make sense to present the analysis of all of it, so I'm just going to summarize what I believe will be of interest and also data that is necessary to illustrate a conclusion.
Just a quick addendum to the previous data. I thought it might be interesting to see if the variations in response time had any correlation to the confidence in the detection result. The correlation is only 2-3% so there really isn't any correlation. Longer detection times do not provide a higher confidence result.
Acquisition Time vs Face Detection Function
The next set of data looks at the acquisition time for the various face detection functions. The minimum detection size is set to 64 pixels and I took the data for a large (F1) and small (F4) version of the same male image that I used before. I tested 6 of the functions separately and then tested a composite of Age/Gender/Expression to test the impact of running several functions simultaneously.
I also ran the test using a large (WF1) and small (WF4) female image to assess the impact of a different image.
It appears that the face detection functions all take about the same amount of average time with some overhead for aggregation of functions. There is a fair amount of variability in the times for all of the functions.
Acquisition Time vs Detection Size
As I've been testing the unit I realize that I've been thinking about the time sensitivity of the detection algorithm incorrectly. I've been assuming that it would take longer to detect a smaller target, but it really is the minimum detection size that drives the acquisition time. That makes sense since the algorithm would have to break the image up into smaller detection boxes. The minimum detection size is a proxy for the maximum range of target detection. In my initial testing the image size didn't affect the acquisition time because I was using the same minimum detection size for all the image sizes.
The smallest minimum detection size that you can set is 20 pixels and the largest maximum is 8192 pixels. The maximum detection size does not have a significant effect on the acquisition time. The acceptable minimum detection size is a function of the maximum detection range. I re-plotted my image size vs distance chart with the distance now in meters.
And I reran the Face and Body Acquisition Time tests with varying minimum detection sizes.
Now I have a good handle on what the necessary acquisition time will be. I want human body detection at 10m and face detection at 2m. I would want a minimum body size of 64 pixels and a minimum face size of 128 pixels. So the body detection time will be 300-400ms and the face detection time will be 200ms.
I'll use those size settings in my application and see if I can get reliable detection at the specified distances.
Top Comments