Ask Your Question

Haar Cascade Classifier

asked 2020-02-14 20:41:14 -0500

Hi, I would like to train my own Haar Cascade and I have the gist of what I need to do but I'm not sure about one thing. So what I'm trying to do is create one for the human body, full body and upper body. If I was to obtain, let's say, 50 images of different humans at different angles and train it, is it going to detect ANY human? Or would it be able to only detect the humans of the images I obtained?

edit retag flag offensive close merge delete


to train a decent cascade classifier, you would need like a few 1000 images, and any variation in pose is harmful. and yes, it will detect any human.

in the end, -- IF you have good data, you should probably train a cnn instead, if you don't, -- it's all a waste of time only.

berak gravatar imageberak ( 2020-02-15 01:21:49 -0500 )edit

Can you define "good data"? Also you said any pose would be harmful, does this mean I can't detect a human crawling, crouching and walking with the same Haar Cascade I trained? Can you elaborate on pose variation being harmful.

dunnoguy gravatar imagedunnoguy ( 2020-02-15 03:00:27 -0500 )edit

I can't detect a human crawling, crouching and walking with the same Haar Cascade


"good data"

e.g. the INRIA person dataset. or CBCL

Can you elaborate on pose variation being harmful.

try to understand the image from the tutorial:

(the bars won't be in the same place, if you allow pose variation)

berak gravatar imageberak ( 2020-02-15 03:39:12 -0500 )edit

see unfortunately it has no source code. but you can try and see possible results

sturkmen gravatar imagesturkmen ( 2020-02-15 05:28:50 -0500 )edit

here you can find a ready LBP cascade for people detection

sturkmen gravatar imagesturkmen ( 2020-02-15 05:32:52 -0500 )edit

I do not want to use a created cascade, I want to create my own, the last thing I'd like to know is, if I train my Haar Cascade with images of persons facing front, would it be harmful to train the same cascade with people facing sideways and facing back? Also if these images are taken directly in front the person, if I angle the camera upwards e.g (cameras on ceilings for surveillance), would it detect the person still? Or do I need to train it at this particular angle? I hope what I'm saying is clear.

dunnoguy gravatar imagedunnoguy ( 2020-02-15 13:37:58 -0500 )edit

1 answer

Sort by ยป oldest newest most voted

answered 2020-03-29 05:13:52 -0500

ComputerVisionary gravatar image

To train a powerful cascade and you can take a look here for a lot of tricks about hyper-speed haar, hog, lbp cascades:

Using 1000 samples is very naive (academic approach), the expert team above uses something like 7-12k positive samples and billions of negative samples with days of training!


edit flag offensive delete link more

Question Tools

1 follower


Asked: 2020-02-14 20:41:14 -0500

Seen: 188 times

Last updated: Mar 29 '20