search

ageitgey / face_recognition

The world's simplest facial recognition api for Python and the command line
MIT License
53.47k stars 13.49k forks source link

face_recognition takes too long to process each frame #884

Open yarin177 opened 5 years ago

yarin177 commented 5 years ago

Ask Question

Hey I am using the following code from face_recognition github: "https://github.com/ageitgey/face_recognition/blob/master/examples/facerec_from_webcam_faster.py" Everything works out beside that I am getting 1 fps/sec when It detects a face, when it is not, the frame rate is ~30.

I looked up on task manager to see how's my computer doing and The CPU is on ~15% as well as the GPU so it isn't hardware problem. I also tried changing the resolution to 320x240 but It did nothing same problem.

EralpB commented 5 years ago

I'd comment out each part and profile, you can save a frame and run a basic loop. Is it the face detection taking time or face matching.

if you have thousands of training pictures maybe comparing each takes time in this way, how many known pictures do you have?

maybe your webcam resolution is very high to begin with, so even resizing it takes time with this method.. ?

yarin177 commented 5 years ago

My webcam resoulution is low and even I tried resizing it to increase the speedwhich did nothing, I think the part that takes long is the face matching

reddytocode commented 5 years ago

using the hog is slow for me, are you using it? or the cnn?, if you have a gpu with cuda using cnn makes it really fast. my options are: 1.- resize your frame just for getting the location of the faces, then crop the faces from the original image (translating your locations to the original size) and with those faces make the recognition. 2.- if you have a gpu but even with the cnn model is slow try using mtcnn for the face location and the dlib for the recognition. 3.- use threads for getting your frame or just dont read all the frames.

i've try with those 3 options and at least one should work for you :D

sandip-narwade commented 5 years ago

@Reddyforcode , I am gonna try with 1 step its a good idea, but I did not understand step 2.

reddytocode commented 5 years ago

@sandipJava i mean even using your graphic card (GPU) the process could be slow, but there's other face detections methods. MTCNN Is a very fast one, faster than the cnn. you can check it here https://www.youtube.com/watch?v=ByV2pj3c4dM

you found a basic implementation on my GitHub.

sandip-narwade commented 5 years ago

@Reddyforcode , what is accuracy of MTCNN campare to CNN

reddytocode commented 5 years ago

@sandip-narwade is not as accurate as CNN but never failed for me in production, have you seen the video?

sandip-narwade commented 5 years ago

@Reddyforcode , I have seen... All are good. Currently, I am working on CNN. If any help I will ping you. Can I send mail to you if I need help?

reddytocode commented 5 years ago

@sandip-narwade Sure write me whenever you want :) but I check often this: readyold@icloud.com

sandip-narwade commented 5 years ago

@Reddyforcode thank you..

congphase commented 5 years ago

using the hog is slow for me, are you using it? or the cnn?, if you have a gpu with cuda using cnn makes it really fast. my options are: 1.- resize your frame just for getting the location of the faces, then crop the faces from the original image (translating your locations to the original size) and with those faces make the recognition. 2.- if you have a gpu but even with the cnn model is slow try using mtcnn for the face location and the dlib for the recognition. 3.- use threads for getting your frame or just dont read all the frames.

i've try with those 3 options and at least one should work for you :D

I don't fully understand your 1st option, does that mean resize the whole input image to a smaller resolution, then apply face detection on that smaller image, then when we get detected face, resize it (only the detected face) to the previous size for recognition?

What if my problem is that I cannot resize the input image to smaller resolution (because faces then are to small to be detected)? Is it okay to keep the image the original size, but crop only the region that faces most likely appear (my application-specific scenario), then apply face detection and then just use it for recognition without resizing anything? Is this okay? Have anyone tried this please give me some advice. Thank you a lot!!!

reddytocode commented 5 years ago

@congphase mtcnn can detect small faces and its faster than cnn and you can pass that face for face recognition

congphase commented 5 years ago

@Reddyforcode Can you share the MTCNN that you've used? I just have a small confusion here, because based on this performance showdown to dlib's cnn, mtcnn, others, mtcnn does not prove to perform better than dlib's cnn.

reddytocode commented 6 months ago

5 Years later: This seems to work just fine https://pypi.org/project/mtcnn/