I was wondering if the Histogram of Oriented Gradients has been made completely obsolete by Convolutional Neural Networks.

HOG descriptors involve specific manually crafted convolutions, while ConvNets learn the best convolutions for the task. Is this proposition generally correct or is there more?

For example, a convolutional layer that may learn edge detections on different orientations (multiple filters) followed by a (learned) Gaussian convolutional layer followed by max-pooling seems to me to achieve a similar result to the HOG descriptor.


I'd say, it's completely application-dependent. ConvNets are designed to serve as a general purpose feature extractor and classifier, but if your domain is simple or your application have no access to GPU ( or is designed to work under really high loads ( like a content filter to detect credit cards labels )) and don't really require all power of convnets -- then you may want not to use ConvNets.

So, let's first clarify your context more precisely

  • 1
    $\begingroup$ for example for binary image classification human vs dog. But is a good ConvNet necessarily more expensive than a HOG+SVM at query time? I don't think so. A HOG+SVM system might even be more expensive, IMHO. $\endgroup$ – fstab Oct 21 '15 at 14:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.