|
|
|
|
|
|
by bdamos
3804 days ago
|
|
|
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. |
|
|
My first problem/observation is that Haar cascades looove running on a GPU due to their Float-y nature. But dealing with them on a CPU frankly stinks. I was getting 1 frame/10 seconds at 800x600 with the included Haar face detector. That's effectively unusable.
Turns out there's also LBP cascades, which are integer based. And they run fast on a cpu. But, from my observations, they have many false positives. But they seem to have no issue with false negatives, so I grab all the faces, plus a few "junks".
The speedup is, is that I can use an LBP and then throw the region of interest (the potential face) onto a Haar cascade eye detector. Now that I'm dealing with much smaller pictures, haar runs acceptably. Literally, if (eyes.size > 0){is valid face.....}
Then I proceed to use the built in function on OpenCV contrib Face library. The problems with the library are numerous. Mainly, the settings are provided without good descriptions, or whatever the defaults of whatever academic papers had them set as.
Because I'm also an academic, I was able to get ahold of quite a few large datasets of face data. After doing so, I wrote a few small programs that attempted to calculate the ideal settings for the FaceRecognizer call, which I believe I did so. (The settings are in the call, in the source.)
Of course, I do get some slowdowns depending on how many faces there are (mainly, stay away from google searches for faces). But then again, 4 haars on 50x50 images is not that bad at all.
My machines used: Thinkpad T61 (8GB ram), Intel NUC (8GB ram, I5 cpu) Camera: Logitech C920 webcam
I did try my code using the max resolution the camera could acquire (1920x1080).... 1 frame/5 seconds.