Deep learning reigns undisputed as the new de-facto method for image classification. However, at least until the beginning of 2016, it was not yet quite clear whether deep learning methods were undisputedly better than more traditional approaches based on handcrafted features for the specific task of video classification. While most research interest is certainly being directed towards deep learning, we thought it would be fun to play on the handcrafted features side for a while and we actually made some pretty interesting discoveries along the way.
Disclaimer: this post is directed towards beginners in computer vision and may contain very loose descriptions or simplifications of some topics in order to bring interest about the field to unexperienced readers – for a succinct and peer-reviewed discussion about the topic, the reader is strongly advised to refer to https://arxiv.org/abs/1608.07138 instead.
Three (non-exhaustive) approaches for classification
In the following paragraphs, we will explore a little three different ways of performing video classification: the task of determining to which of a finite set of labels a video must belong to.
Handcrafted features: designing image features yourself
Initial video classification systems were based on handcrafted features, or, in other words, based on techniques for extracting useful information from a video that have been discovered or invented based on the expertise of the researcher.
For example, a very basic approach would be to consider that whatever could be considered as a “corner” in an image (see link for an example) would be quite relevant for determining what was inside this image. We could then present this collection of points (which would be of much reduced size than the image itself) to a classifier and hope it would be able to determine its contents from this simplified representation of the data.
In the case of video, the most successful example of those is certainly the Dense Trajectories approach of Wang et al., that captures frame-level descriptors over pixel trajectories determined by optical flow.
Deep learning: letting networks figure out features for you
Well, some could actually think that using corners or other features to try to guess what is in an image or video is not straightforward at all – why not simply use the image or video itself as a whole and present this to the classifier to check out what it would come up with? Well the truth is that this had been tried, but until a few years ago, it simply wouldn’t work except for simple problems. For some time we didn’t know how to make classifiers that could handle extremely large amounts of data, and yet extract anything useful from it.
This started to change in 2006, when the interest on training large neural networks started to raise. In 2012, a deep convolutional network with 8 layers managed to win the ImageNet challenge, far outperforming other approaches based on Fisher Vector encodings of handcrafted features. Since then, many works have shown how deep nets could perform way better than other methods in image classification and other domains.
However, while deep nets have certainly been shown to be the undisputed winners in action classification, the same could not yet be said about video classification, at least not in the beginning of 2016. Moreover, training deep neural networks for video can also be extremely costly, both in terms of computational power needed as well as the number and sheer size of examples needed to train those networks.
Hybrid models: getting the best out of the both worlds 😉
So, could we design classification architectures that could be both powerful and easy to train? In the sense that it wouldn’t be necessary to rely on huge amounts of labelled data in order to learn even the most basic, pixel-level aspects of a video, but instead in a way that we could leverage this knowledge already from techniques that are already known to work fairly well?
Well, indeed, the answer seems to be yes – as long as you pay attention to some to some details that can actually make a huge difference.
This is shown in the paper “Sympathy for Details: Hybrid Classification Architectures for Action Recognition.” This paper is mostly centered about two things: a) showing that it is possible to make standard methods perform as good as deep nets for video; and b) showing that by combining Fisher Vectors of traditional handcrafted video features with a multi-layer architecture can actually perform better than most deep nets for video, achieving state-of-the-art results at the date of submission.
The paper, co-written with colleagues and advisors from Xerox Research Centre and the Computer Vision Center of the Autonomous University of Barcelona, has been accepted for publication in the 14th European Conference on Computer Vision (ECCV’16), to be held in Amsterdam, The Netherlands, this October, and can be found here.