Can I decrease the recognition time - YOLO

Question

My Jevois camera/ Arduino Nano combination alerts me wonderfully when it sees a "person", "motorbike", or  "bicycle". The acquisition time is adequate but I would like for it to be even faster. Currently, I'm running " YUYV 640 480 15.0 JeVois DarknetYOLO ".  Is this the best choice for person recognition? Please advise, Thank you!!

Answer 1

great to hear! Yes, you can choose speed based on resolution. See the section about "speed and network size" at http://jevois.org/moddoc/DarknetYOLO/modinfo.html

so, if you do, for example,

setpar netin 160 120

you should process at about 695ms/frame (1.4 frames/s). You could also reduce to YUYV 320 240 15.0 in such case. Note that grabbing twice bigger than you will process is desirable as the downscaling by a factor 2 (here, grab at 320x240, then scale down to netin at 160x120) will cleanup some YUYV artifacts (color information is low resolution in this format which only uses 2 bytes/pixel).

also see the second half of this video: https://www.youtube.com/watch?v=77VRwFtIe8I

the faster you get, the less accurate you will be.

Comments

Thank you! Will try shortly.