Missing HOG Cascade in Ver 3

edit

They mentioned that "The implemented HOG features are quite weird - different from Dalal's interpretation of HOG, different from P. Dollar integral channel features"

I don't understand why it should be omitted if it is not the same as Dalal's interpretation!!! It was working very good and in many cases the results were better than Haar cascade and LBP!!

Actually I think dropping faulty implementations is a good approach since you are telling people that it follows a certain convention while it does not. It will return once somebody can integrate the working HOGDescriptor interface into this. In the meanwhile, just use the 2.4 branch if you want this or port the functionality locally.

I think instead of dropping, the author of the code can write a document and describe the method.

The original author was not available for help anymore as I understood from one of the core developers. And actually, there is only 1 way of calculating HOG features so we cannot just add a semiHOG implementation because it does something good... that is not the purpose of a computer vision library...

It is strange that the HOG implementation was not documented so it may be rewritten... I suggest you to do documentation of all the approaches in OpenCV for not having the same problems later... ;)

so.. it is not documented anymore but is still callable? because im pretty sure im calling it and getting back a classifier that seems to work decently in certain params and situations (opencv 3 python API)

@jsky there is a difference between HOG cascades and HOG+SVM as the HOGPeopleDetector. The latter still works perfectly fine! The first one is disabled.

but i call pkg-config --modversion opencv : 3.0.0 and opencv_traincascade ... -featureType HOG and i get back a cascade that seems to use HOG features <featureType>HOG</featureType>.. Are you saying this is using an SVM implementation as opposed to decision tree?

Well that is because the traincascade app has not been adapted according to the detection algorithm. So you will be perfectly fine in training one, but as I recall it, at calling detectMultiScale, the featureEvaluator does not know how to handle the HOG features anymore.

• As you can see here, in the latest master branch, there is no line anymore to evaluate the HOG features.
• Checked it for 3.0.0 also and there it is also missing.

So even if you train a model using the faulty feature set, you won't be able to call it at detection time.

hmm i see.. but it seems that i am calling it in detection and getting decent results. i guess its because its still in the python api? as im using python. nowhere do i tell it to use a particular evaluator, i just tell it about the classifier and then call detectMultiscale ..

How are you calling it? We are not talking about the same interface, else it would NOT work, because the python api is a simple wrapper of the C++ API ...Start by giving us some more info.

my bad. i am using cv2 in detection but cv3 in training (as above).

Then indeed, 2.4 should still support it, where 3 does not. We should remove it from the cv3 interface from trainng also. Put it on my weekly todo list!

It seems that HOGCascade implementation in opencv2.4 is similar to this approach: "Fast Human Detection Using a Cascade of Histograms of Oriented Gradients" http://lear.inrialpes.fr/people/trigg...

So, it seems that in contrast to what @Dikay900 and @vpisarev mentioned, it is a valid implementation.