bdamos

Thanks for the feedback, I just updated the animation to pause at the beginning and end to make the completions clearer: https://github.com/bamos/bamos.github.io/commit/f6a152851355...

The argument isn't that you use OpenCV: OpenFace also uses OpenCV. However, I think you should target and present your program as being a program that uses face recognition, not as a face recognition program. You are using and not crediting here that your program uses existing, off-the-shelf face recognition functionality already in OpenCV: https://github.com/jwcrawley/uWho/blob/2823479d5abf9f8f2de21...

OpenFace can optionally use a CUDA-enabled GPU, but it's not a requirement. The performance is almost real-time on a CPU. After detection (which varies depending on the input image size), the recognition takes less than a second. We have a few performance results on the FAQ at http://cmusatyalab.github.io/openface/faq/

I'm surprised (and skeptical) uWho can do detection+recognition at 15fps. I would expect face detection alone in 1280x720 images to be much slower than 15fps. On my 3.7GHz CPU with a 1050x1400px image, dlib's face detector takes about a second to run. This is also my experience with OpenCV's face detector, which I noticed your code is using. Also OpenCV's face detector returns many false positives, especially in videos. See this YouTube video for an experimental comparison: https://www.youtube.com/watch?v=LsK0hzcEyHI

Also, I think it's a strong claim that faces can't be generated from a perceptual hash. One property of perceptual hashes is that hashes that have a close hamming distance to each other are more similar (of the same person). I wouldn't be surprised if a model could successfully map perceptual hashes to faces given enough training data. I read a good paper about doing this (not specific to faces) but can't remember the reference now.

Edit: I just added some simple timing code to this sample OpenCV face detection project on my 3.60GHz machine: https://github.com/shantnu/FaceDetect On the John Lennon image from the OpenFace FAQ sized 1050x1400px, it takes 0.32 seconds, which is about 3fps. This is slightly quicker that dlib's detector on the same image, but it also returned a false positive.

Thanks for pointing out the 404, I just corrected the link.

There's an interesting discussion on lobste.rs from a few months ago about privacy issues and licensing: https://lobste.rs/s/sajz0s/openface_face_recognition_with_go....

Summary: OpenFace uses fundamentally different techniques (a deep neural network) for face recognition that OpenBR currently doesn't provide.

--

As our initial ROC curve on LFW's similarity benchmark in https://github.com/cmusatyalab/openface/blob/master/images/n... shows, this approach results in slightly improved performance. The best point is an FPR of 0.0 and TPR of 1.0 (top left). You can see today's state-of-the-art private systems in the top left, followed by open source systems, then by historical techniques OpenCV provides like Eigenfaces. The dashed line in the middle shows what randomly guessing would provide.

OpenBR is going in a great direction for reproducible and open face recognition. They provide a pipeline for preprocessing and representing faces, as well as doing similarity and classification tasks on the representations. The techniques from OpenFace could be integrated into OpenBR's pipeline.

Thanks for the offer! Our original model `nn4.v1` should perform OK on your data if you're interested in trying to automatically predict people in new images.

Training new models is currently dominated by huge industry datasets, which currently have 100's of millions of images. My current dataset is from datasets available for research and has ~500k images.

Yes, the processing pipeline first does face detection and a simple transformation to normalize all faces to 96x96 RGB pixels. Then each face is passed into the neural network to get a 128 dimensional representation on the unit hypersphere.

For a landscape, face detection would probably not find any faces and the neural network wouldn't be called.

And an image with multiple people will have many outputs: the bounding boxes of faces and associated representations.

This depends on what you want to use face recognition for. Maybe I should say more clearly in the README who this project is for. I could have released trained classifiers on 10,000 celebrities, but I focused the projects towards providing an easy way to train new classifiers with small amounts of data. I think this direction allows for more people to use and benefit from the library.

For example, check out our YouTube video of a demo training a classifier in real-time with just 10 images per person at https://www.youtube.com/watch?v=LZJOTRkjZA4. This demo is included in the repo and the README has instructions on running it.

Also note that there is a distinction between training the neural network, which extracts the face representations, from using the features for tasks like clustering and classification.

Thanks, I'm hosting on a nonstandard port on a server. I put the main page inside a frame hosted on GitHub pages for a better URL, but this broke the links. Fixed now. :-)

I have two minor comments:

1) Linking to the original HN hiring post is helpful, and

2) I'm on a poor internet connection now, and this page took about a minute to load. http://hnhiring.com/ and the original July 2015 thread (https://news.ycombinator.com/item?id=9812245) load in ~5 seconds.

How does the speed compare to other compression algorithms?

Another reason for me is the cross-platform compatibility. With git, I'm able to synchronize passwords across OSX and Linux machines, and the features (even copying) work well on both platforms.

How did you select words to compare? Did you have to try many poor combinations before selecting a "good" set?

This year's CVPR has a lot of great papers. In case anybody hasn't seen, a full list's available at http://cs.stanford.edu/people/karpathy/cvpr2015papers/

Shameless plug for one of my side projects to manage reading lists with YAML and GitHub pages: https://github.com/bamos/reading-list