Hacker News Reader

Comment by declan | Hacker News Reader

Whether auto-face recognition is a good idea or not, it's going to happen as the trends we all know and love continue. Even if major tech companies don't build this in to their camera-enabled products, open source projects will using distributed datasets. Or perhaps not-so-distributed datasets: face recognition systems use only something like 70-80 nodal points that give you internodal distances like eye spacing, mouth width, etc. (I may be wrong on that number; if I am I'm sure someone who works in this area will correct me.)

If this happens, I suspect it will change human interaction significantly and make it more different to be casually anonymous. I'm not so worried about people I casually interact learning my identity; I'd be more concerned about cradle-to-grave government collection and permanent storage of these records, coupled with license plate scanners, etc.

View on HN

Yea, so at Lambda, depending on the application, we use either convolutional neural networks or layered neural networks initialized with unsupervised training using autoencoders to do the feature extraction from images. The features are either passed through a softmax layer or used as a "siamese" network. (Siamese networks provide a similarity metric as opposed to a probability distribution over labels.[3])

As far as recognition goes, most modern algorithms don't use fiducials (what you call 'nodal points') to do recognition. So there's nothing in our code that has to do with interocular distance or mouth width, etc., so all of the features are "learned". However, fiducials are used to do a frontalization or re-alignment step before passing it through the feature extractor (neural network or otherwise). This type of frontalization step allowed Facebook's deepface algorithm to use locally learned filters for their CNN as opposed to shared weights.[1][2]

[1] http://www.cv-foundation.org/openaccess/content_cvpr_2014/pa...

[2] http://mmlab.ie.cuhk.edu.hk/pdf/YiSun_CVPR14.pdf

[3] http://yann.lecun.com/exdb/publis/pdf/chopra-05.pdf

I'm a research scientist at a face recognition company. From my experience, we don't use internodal distances, we use information from all the pixels.

Would you be so kind as to expand slightly on what "information from all the pixels" means to us plebeians?

Would bet state-of-the-art is doing face recognition with http://en.wikipedia.org/wiki/Convolutional_neural_network or similar.

Well with regards to the post I answered, some have the idea that you locate certain features like eye corners, mouth corners, etc. and then do recognition based on the measurements between these feature locations.

One of the earlier successes of face recognition as described in Pentland and Moghaddam's paper was to use simple PCA on the images of faces normalized by eye position. This PCA would use all the pixels of a vector made from an image. Obviously there's been a lot of progress since then, but all pixels are getting used for information.

A couple years ago I worked on a product that added a feature to do, not exactly facial recognition, but "facial similarity" perhaps. Not any kind of bleeding-edge research but just a helpful little value-add. The approach we took was exactly the kind of geometrical measuring of nodes that I believe you are referring to. It worked okay.

Principal Component Analysis looks interesting, but I'm struggling to understand how 2D pixel data would be transformed into the appropriate covariance matrix. Any insight on the missing link you could offer?

I would look here: http://en.wikipedia.org/wiki/Eigenface

The technology has advanced far beyond that, but that was the start where automatic face recognition first became realistically possible.

Reply on news.ycombinator.com