Train from haarcascade frontface

asked 2020-06-26 11:08:43 -0600

Hi, this must be a question that is often asked, but I cannot find an answer nor intructions on how to.

Is it possible to train a cascade model from the already train haarcascade frontface default? My goal is for the script to recognise specific people (i.e. me). I know there is a recongniser built in opencv, but it is quite slow on a raspberry pi compared to a cascade xml.

I am tryingto train a model from scratch, but with mixed results.

thanks

(ps: I'll also post this question on stackoverflow)

edit retag flag offensive close merge delete

Comments

I know there is a recongniser built in opencv, but it is quite slow on a raspberry pi

you must be doing something wrong

berak ( 2020-06-26 11:26:44 -0600 )edit

Really? what fps should I be expected? I got barely a couple.. I followed this post https://towardsdatascience.com/real-t...

willmendil ( 2020-06-26 12:12:38 -0600 )edit

how large are your crops ? most time will be spent in feature extraction

if those are significantly larger than say100x100 -- resize to some thing smaller

I got barely a couple..

yea, but you also got those weird sleep() and (blocking) GPIO calls in there, once you detect something, so no fair measuring possible

berak ( 2020-06-26 12:22:29 -0600 )edit

humm, interesting, much better frame rate, now I just need to deal with the latency.

willmendil ( 2020-06-26 12:33:43 -0600 )edit

add a comment

answered 2020-06-26 11:18:28 -0600

berak
32993 ●7 ●81 ●312

updated 2020-06-26 11:46:50 -0600

no, you cannot do that.

while you can stop and restart your own training, and it will restart at the last stage computed before, you will need the whole original data (the "vec" file, remember ?) and all the previous (temporary) stage cascades

you don't have this data, so -- no way.

I am tryingto train a model from scratch, but with mixed results.

yea, you need tons of positives for this, might need 5k+ to get better than the prebuilt cascades

My goal is for the script to recognise specific people

no, that won't work, either,

you need face recognition (who is it?) and this does face detection (face or not, and where ?)

(you probably try to train a cascade per person, just forget it, it's proven not to work)

Comments

that's what I don't understand in the lexicon, what do you mean be "prebuilt cascades"

willmendil ( 2020-06-26 11:30:43 -0600 )edit

prebuilt cascades

like the haarcascade_frontalface.xml that comes with opencv

berak ( 2020-06-26 11:43:39 -0600 )edit

"(you probably try to train a cascade per person, just forget it, it's proven not to work)" Yep that was exactly what I had in mind. Thanks for the heads-up. Any references? ok I see what you mean by prebuilt. Thanks very much, that was very insightful

willmendil ( 2020-06-26 12:09:33 -0600 )edit

add a comment

Train from haarcascade frontface

Comments

1 answer

Comments

Links

Question Tools

Stats

Related questions

Train from haarcascade frontface edit

Comments

1 answer

Comments

Links

Question Tools

Stats

Related questions

Train from haarcascade frontface