Ask Your Question

Haar classifier limitations

asked 2016-10-30 03:08:01 -0500

Mohamed Alzaki gravatar image

updated 2016-10-30 17:47:54 -0500

I am working on a project that should take a ct scan image as an input and it should detect the kidney and do some processing on it, my question is due to the low features in that type of images and variant types of kidneys, is haar classifier qualified to detect such object?

The objects in the red rectangles (kidneys) are the target objects which i want to detect.

image description

Another question i am training with the samples at the time i am writing this post, with 579 positive images and around 1k negative images i am using haartraining and it is running for 14 hours now and is in stage 4 but i've noticed in that stage that the SMP is always around 0% - 2% is that normal?

image description

Here is the parameters i use for haartraining

haartraining.exe -data cascades -vec vector/kidney.vec -bg negative/bg.txt -npos 579 -nneg 973 -nstages 15 -mem 1024 -mode ALL -w 24 -h 24

rem -nonsym

edit retag flag offensive close merge delete


a small example img might be helpful

berak gravatar imageberak ( 2016-10-30 03:31:26 -0500 )edit

@berak updated please check the images thanks for your note.

Mohamed Alzaki gravatar imageMohamed Alzaki ( 2016-10-30 05:04:27 -0500 )edit

@StevenPuttemans i am very interested to hear your opinion in this matter.

Mohamed Alzaki gravatar imageMohamed Alzaki ( 2016-10-30 05:22:05 -0500 )edit

24 hours passed in the same stage with exactly same SMP and same HR but the FA decreased a little bit I've updated the answer with the command i use for haar training for the second matter.

Mohamed Alzaki gravatar imageMohamed Alzaki ( 2016-10-30 17:40:43 -0500 )edit

I guess its hanged there because HitRate doesn't increase at all for more than a day now but FalseAlarm does change, i think its maybe because i didn't set sufficient parameters? or i am wrong, in some cases it takes too long with the same numbers.

Mohamed Alzaki gravatar imageMohamed Alzaki ( 2016-10-31 02:00:02 -0500 )edit

1 answer

Sort by ยป oldest newest most voted

answered 2016-10-31 06:00:35 -0500

Ok lets get you some answers

  1. I am unsure, if cascade classifiers is actually what you want, since I am unsure how much variation a kidney object can have. But it seems to be a suitable approach.
  2. That being said. drop haartraining completely. It is old, deprecated, not supported anymore and indeed runs for ages without improving.
  3. Instead, use the newer traincascade interface, which is recent, fully supported and has tons of bugfixes compared to the haartraining tool.
  4. To reduce training time, first build a classifier with LBP features, and once you have decent results, you can let your system train a HAAR model. But LBP is better for parameter tuning for your case, since it trains way faster.
edit flag offensive delete link more

Question Tools

1 follower


Asked: 2016-10-30 03:08:01 -0500

Seen: 2,145 times

Last updated: Oct 31 '16