In part 1 and part 2 I discussed a paper on face recognition and the methods it relies on. Recall, each 100×100 gray scale image is a table of 100×100 = 10,000 numbers that can be rearranged into a 10,000-vector or a point in the 10,000-dimensional Euclidean space. As we discovered in part 2, using the closedness of these points as a measurement of similarity between images ignores the way the pixels are attached to each other. A deeper problem is that unless the two images are aligned first, there is no way to use this representation to discover that they depict the same or similar thing. The proper term for this alignment is image registration.

The similarity between images represented this way will be entirely based on their overlap. As result, the distance can be large even between images that we would consider similar. In part 2 we had examples of one-pixel images. More realistic examples are these:

• image with an object in one corner onewith the same object in another corner;
• image of a cross and the same cross turned 45 degrees;
• etc.

Back to face identification. As the faces are points in the 10,000-dimensional space, these points should be grouped somehow. The point is that all images of the same individual should belong to one group and not any other. It is common to consider “clusters” of points, i.e., groups formed of point close to each other. This was discussed above.

Now, in this paper the approach is different: a new point (the face to be identified) is represented as a linear combination of all other points (all faces in the collection).

As we know from linear algebra, this implies the following. (1) the entire collection has to be linearly dependent, (2) you can find a subcollection that adds up to 0! In other words, everything cancels out and you end up with a blank photo. Is it possible? If the dimension is low or the collection is large (the images are small relative to the number of images), maybe. What if the collection is small? (It is small – see below.) It seems unlikely. Why do I think so? Consider this very extreme case: you may need the negative for each face to cancel it: same shape with dark vs. light hair, skin, eyes, teeth (!).…

Second, the new image in the collection has to be a linear combination of training images of the same person. In other words, any image of person A is represented as a linear combination of other images of A in the collection, ideally. (More likely this image is supposed to be closer to the linear space spanned by these images.) The approach could only work under the assumption that people are linearly independent:

No face in the collection can be represented as a linear combination of the rest of the faces.

It’s a bold assumption.

If it is true, then the challenge is to make the algorithm efficient enough. The idea is that you don’t need all of those pixels/features and they in fact could be random. That must be the point of the paper.

The testing was done on two collections with several thousand images each. That sounds OK, but the number of individuals in these collections was 38 and 114!

To summarize, there is nothing wrong with the theory but its assumptions are unproven and the results are untested.

P.S. It’s strange but after so many years computer vision still looks like an academic discipline and not an industry.