3 ms·
This is rather neat, though the actual paper is less breathless than the wired article. I skimmed through it, and here's what I got: The algorithm assumes fac
by iamwil 19y ago
This is rather neat, though the actual paper is less breathless than the wired article. I skimmed through it, and here's what I got:
The algorithm assumes face images have been cropped and normalized, and only has been run on frontal views. Pose variation, detection, and alignment aren't handled by this algorithm--claimed to be a different problem. But it does handle different illumination and expressions, as well as occlusion and noise.
According to No-free-lunch theorem, there is no generalized classifier that performs better than any other classifier on all classification tasks. Therefore, in any classification task, the solution is often tailored towards the problem, and picking features is one way to tailor it. Often times, you can't use all the features available to you, because they could be noisy or using them all can be computationally expensive. This is why people use SVD, PCA, or other techniques to filter out features that affect the classification the least, so you only do classification on features that affect classification the most.
That's why people use some heuristics like picking eyes, mouth, noses as features, as it would seem like that's where most of the information is.
The weird thing being claimed here is that it actually doesn't matter which features you pick as long as 1) feature space is huge 2) you pick enough features 3) you use a "correct" sparse representation of the features
Apparently, they do recognition by trying to represent the new face as a linear combination of all known faces in the database. So if the new face belongs to a person in the database, it can be represented mostly by faces in the database known to belong to one person. If the new face belongs to someone not in the database, then we can only represent it as a combination of different peoples' faces in our database.
In order for the surveillance mentioned in Wired to work, you'd still need to have an annotated database of people's faces built from somewhere. Same with auto-tagging people's photos on facebook. And you'd need more than one photo of a person too.
However, given Louis Van Ahn's method of using games to entice people to do recognition tasks, and facebook tagging, you can create database of people's faces fairly easily.
- pixcavator 19y ago>The weird thing being claimed here is that it actually doesn't matter which features you pick as long as 1) feature space is huge 2) you pick enough features 3) you use a "correct" sparse representation of the features I haven't read the paper yet, but from your description this part is indeed the least convincing. It still matters which features you use! What if I pick the average color of the image (and millions of similar "features")? A feature like that won't help me in face recognition.
- iamwil 19y agoYou're in computer vision, so you'd know more than me. Like you, I am a bit skeptical about it, but maybe this particular technique happens to work well enough for faces, due to the nature of face image data. Just like Bayesian classifiers assume independence when there isn't any, it works "pretty damn well". If you do get around to reading it, lemme know if you picked anything up from it that also smell fishy.
- pixcavator 19y agoWhat they say "features" they mean simply pixels! I have no idea why. So, if you have a collection 100x100 images, they are all points in a 10,000 space. This seems to be a common approach in pattern recognition. Now for face identification, 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 collection. 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.
- pixcavator 19y agoThe testing was done on two collections with several thousand images each. But the number of individuals in these collections was 38 and 114!
- iamwil 19y agoYeah, they simply mean pixels. I failed to mention that in my previous post. It seems odd to me that it works because doesn't representing it as a linear combo of pixels mean they throw out structural information of the image? I don't know how they came to make that assumptions, but could it be that the linear independence comes about because of the high dimensionality? The higher the dimensions, the harder it is to find a combination of sparse representations (linear comb) that gives you the sparse representation matching the test image. And if you do, and it all comes from the same person, then they say it matches. There must be some shape or property of the face subspace that allows for this, since potentially, you can match a duck to someone's face.
- aswanson 19y agothey do recognition by trying to represent the new face as a linear combination of all known faces in the database What is the difference between doing this and just finding the dot product of the facial feature eigenvectors with the feature eigenvalues and choosing the maximal projection? That's how it's been done since forever, AFAIK.