I know there is already HOG descriptors + linear SVM method for human detection. Is there any way to improve it, maybe by using RCNN or such? The one that I can explore more (for object detection like cars).
For example, from the link below, two different answers (can and can't) are present for the use of CNN with HOG descriptors, and for a beginner like me, their arguments are kind of confusing. I hope someone can help me.
Thanks in advance.
https://www.quora.com/How-do-I-use-deep-CNN-to-learn-HOG-descriptors
Hi,
Interesting question, I have recently started studying deep learning and its application to computer vision, one thing which is pointed' out that the deep networks do not require 'hand-crafted' features. Hidden layers take care of it. So you can train your CNN on the image dataset and then compare it with SVM+HOG.
Take a look at the attached file, it is a transcript of Q&A session. The first question is really interesting
You can use HOG as the input of CNN, but it is not common to do so. In most cases, it under-performs pure CNN.
Note that CNN does not equal R-CNN. The latter is based on selective search (image segmentation), and so the input has to be an image (you cannot use HOG as the input of R-CNN).
R-CNN is already an old solution for object detection. The newest one is called Faster R-CNN. It is much faster than R-CNN, and in most cases more accurate than R-CNN. Still, you cannot use HOG as the input here. Please check the following link:
https://github.com/rbgirshick/py-faster-rcnn
Note that there is a method called Fast R-CNN, which is an older version of Faster R-CNN.
- Badges
- Science topic