Ask Your Question
0

Multiple object detection using HOG+SVM?

asked 2017-04-30 04:29:45 -0500

John Strood gravatar image

I read a lot of posts on the web that said it was impossible to detect objects of different types like human face, car, dog,cat, banana etc using a single classifier. But I began wondering can we not train a classifier with images of different objects in a multiclass SVM? The SVM predict then returns the class label from which we can infer the type of object. I know this is really simple task by using a Caffe's RCNN. But I wished to know if the method I described was feasible.

edit retag flag offensive close merge delete

1 answer

Sort by ยป oldest newest most voted
3

answered 2017-04-30 06:08:15 -0500

berak gravatar image

updated 2017-04-30 06:13:06 -0500

  • using opencv's HOGDescriptor, you can only detect objects of a single class. (it's using a single, binary regression support vector in this case)

  • but you can also compute your own HOG features, and use those with multi-class SVM for classification.

if you're using an SVM, you simply cannot to both at the same time, with the same setup. (it's a bit different with rcnn or yolo, where you get both segmentation and identification)

edit flag offensive delete link more

Comments

Yes, I was indeed talking about the second possibility. That is compute the HOG feature vector and perhaps pair it with LibSVM. But no one's ever talked about it. So I was thinking if it were ever feasible.

John Strood gravatar imageJohn Strood ( 2017-04-30 13:13:06 -0500 )edit

Hi, where you able to find a solution for the same issue? I was also trying to do multiple object class recognition using image features , however not getting any leads on the same? Please let me know if you have a clue to this. Mostly I find codes only for a single object class (like either cars or humans, but not both)

chandini gravatar imagechandini ( 2019-02-11 21:24:34 -0500 )edit

@chandini -- what's unclear about the answer above ? please try to read it again.

berak gravatar imageberak ( 2019-02-12 00:22:51 -0500 )edit

I am clear about the above answer, I was asking whether it was possible to compute our own HOG feature vector. And if yes, any leads on how they did it?

chandini gravatar imagechandini ( 2019-02-12 09:21:25 -0500 )edit

I think i saw your comment from my gmail notification for the below question i posted on the same: http://answers.opencv.org/question/20...

It seems convincing... However, I find you deleted the same. :( Any other suggestions? @berak

chandini gravatar imagechandini ( 2019-02-12 09:29:58 -0500 )edit

there is code to train HogDescriptors for detection here but again -- for a single class (vs "background") only

berak gravatar imageberak ( 2019-02-12 09:33:59 -0500 )edit

@berak... so if i want to detect or match both cars and humans I probably need two different codes right? That is one for cars and then one for humans.?

chandini gravatar imagechandini ( 2019-02-12 09:35:36 -0500 )edit

yes, unfortunately. if you also need bycicles & wheelchairs, you're already at 4 detectors.

i think you should give up your deep learning aversion. resistance is futile here ;)

you could simply pickup something pretrained from here

berak gravatar imageberak ( 2019-02-12 10:08:13 -0500 )edit

ha ha sure. I was trying a different approach from deep learning. :)

chandini gravatar imagechandini ( 2019-02-12 10:09:16 -0500 )edit
Login/Signup to Answer

Question Tools

2 followers

Stats

Asked: 2017-04-30 04:29:45 -0500

Seen: 1,370 times

Last updated: Apr 30 '17