Open RayLuxembourg opened 6 years ago
The camera I am testing with has a max resolution of 1900x1080 but I configured the HTML camera object's resolution to 1280x720, and my monitor is 1080 native. I find that it stops detecting faces below 39x39 px so it may be dependent on your monitor resolution or the resolution of the camera.
@alexandermaisey could you share your settings?
getUserMedia({ audio: false, video: { facingMode: 'user', width: { ideal: 1280 }, height: { ideal: 720 }, frameRate: { ideal: 10, max: 10 } }})
And my setEdgesDensity is set to 0.2, everything else is default
@alexandermaisey
tracker.setInitialScale(4);
tracker.setStepSize(2);
this what you mean default?
I haven't defined initialScale or stepSize, so they go to their default values; I believe stepSize is 1.5 and initialScale is 1.0 TBH I didn't understand what they did so I decided it was best to not mess with them.
@alexandermaisey something is very wrong with the results..
I tried setting tracker.setInitialScale(4) and tracker.setStepSize(2) and my application basically stopped detecting faces below or above 100px
When I modified to setEdgesDensity(0.1) setInitialScale(2) and setStepSize(1.2) it gets less weird and detects more size faces but still nothing under 39px
Like I said, I'm not sure what these do. Any clues?
@alexandermaisey the documentation did poor job explaining this properties so I cant be sure. all i know is that i am able to detect down to 39px max but accuracy is weird.
In order to reduce detecting time, I have to set a relatively small size for the images, basically I set the size nearly as the face(image) example given on the official website.
I used stepsize 1.7, which is used by the example, and keep everything else as default. But I got really low accuracy for face detection. The algorithm sometimes detect faces on images with no faces actually and sometimes it cannot detect any faces when there actually are. The estimate accuracy rate is like 50%, unacceptable.
I tried several combinations of image sizes/stepsize, apparently the image size affects accuracy greatly but I haven’t found a proper size/stepsize yer. Should I really mess with scalefactor and edgedensity? I don’t have a clue which value combinations should be given instead of the defaults.
Where is explained the settings for:
Can someone explain a proper configuration ?
Thanks.
Is it possible to detect faces with rect width and height of around 50px?