I'm surprised this works well. If the method is a state-of-the-art convolutional neural network architecture, you generally would need more than 10 images -- even with transfer learning -- for passable accuracy. Algorithms for medical diagnosis, for example, generally require between 100 and 200 images to do well. Though those are generally transfer learned from ImageNet CNNs. So I'm curious as to which dataset this face recognition uses for weight initialization, or it uses another ML method entirely.
Well before deep learning you could get good performance on people in your training set. Source worked as research scientist for a face recognition company since 2003.
Pre deep learning -- using facial feature geometry?
What was the use case? Was it for determining faces in general or distinguishing between different faces?
I used a convolutional neural net for face recognition on a hobby project, but I kept getting issues where the probability of matching a face was high so long as a face was actually in the image. Unfortunately, this didnt work well with a sliding window algo because Id get a bunch of windows with a high probability of a face with the only difference being a slight shift in increment / size. Would it be better to just use a multi layer perceptron? Also, does anyone else find it amusing when their face recognition systems identify things like toes or gates with high confidence? I end up spending some time zooming in to make sure it's not one of those hidden faces in random images
How much time did you spend before you felt stuck?
I've probably invested 40 hr in the project, but half of it was spent on iterating through bad examples, and identifying them as negative labels for training data. I also found that the open source data sets tend to have full portraits instead of just the face which took some more data clean-up. My end goal is just to have the system be able to detect faces in a video stream, and have the camera follow you around, but I'm not comfortable moving on until the false positive rate goes down. It's not a huge problem for my use case, but I was hoping to be able to detect all faces in any size pic regardless of how far a face is in the image. I find ML development to be a little more annoying because sometimes tuning the hyper parameters can feel like magic as opposed to actually learning something. I kinda want to go back to the data and get rid of all the portrait style pics. Thinking about that as a weekend project doesn't exactly get me up in the morning though XD. Probably worth going through the fast AI course too because my ML experience ends with the ml course from Andre Ng