Face not detected in an image: number_of_times_to_upsample=2 results in exception

[pmi123]

• face_recognition version: face-recognition==1.2.3, face-recognition-models==0.3.0
• Python version: 3.4.3
• Operating System: Ubuntu 14.04
• Django 2.0.5
• Celery 4.2.1
• Redis 2.10.6

Description

My python program detects faces in many photographs (not from cell phone, but scanned images). The baby's face is not detected in one photo (052_002), but was detected in a similar photo (052_003) (both are attached). All the other faces were detected in both photos (including the baby in 052_003). 052_002 - baby looking at mother 052_003 - baby looking at camera

What I Did

When the baby's face in 052_002 was not found, I looked through your web site and found this article - https://github.com/ageitgey/face_recognition/wiki/Face-Detection-Problems, so I tried adding the parameter "number_of_times_to_upsample=2" in face_encodings, but got this error: face_encodings() got an unexpected keyword argument 'number_of_times_to_upsample'

Paste the command(s) you ran and the output.

image_path = os.path.join(document_path, document_file_name) image = face_recognition.load_image_file(image_path) face_encodings = face_recognition.face_encodings(image, number_of_times_to_upsample=2)

If there was a crash, please include the traceback here.

[2018-11-12 11:06:04,077: ERROR/ForkPoolWorker-1] Task biometric_identification.tasks.find_faces_task[dd9be9ee-6529-4a1c-8327-332a5b0b9532] raised unexpected: TypeError("face_encodings() got an unexpected keyword argument 'number_of_times_to_upsample'",) Traceback (most recent call last): File "/home/mark/.virtualenvs/memorabilia-JSON/lib/python3.4/site-packages/celery/app/trace.py", line 382, in trace_task R = retval = fun(*args, *kwargs) File "/home/mark/.virtualenvs/memorabilia-JSON/lib/python3.4/site-packages/celery/app/trace.py", line 641, in __protected_call__ return self.run(args, **kwargs) File "/home/mark/python-projects/memorabilia-JSON/biometric_identification/tasks.py", line 55, in find_faces_task face_encodings = face_recognition.face_encodings(image, number_of_times_to_upsample=2) TypeError: face_encodings() got an unexpected keyword argument 'number_of_times_to_upsample'

Is it possible to change the parameters for face_encodings to find this face?

Is there a way to select the face along the lines of Facebook tagging, and then get the location and encodings so I can add them to my database?

Thanks!

Mark

[ageitgey]

The parameter number_of_times_to_upsample is for the face_locations() function, not the face_encodings() function. To use it, you need to break out face detection as a separate step. Try changing your code to this:

image_path = os.path.join(document_path, document_file_name)
image = face_recognition.load_image_file(image_path)
face_locations = face_recognition.face_locations(image, number_of_times_to_upsample=2)
face_encodings = face_recognition.face_encodings(image, known_face_locations=face_locations)

[ageitgey]

And by splitting it into two steps like this, this also gives you an array of face locations that you could save in a database (so you know which face is which)

[pmi123]

Thanks!! I totally missed that very important point about face_locations versus face_encodings.

My code does look just like yours, and I do save the locations in the database along with the encodings to make recognizing the faces easier. I just left out those details in my first question.

Unfortunately, making both the changes you described still did not find the baby's face in image 052_002.

Any examples of how to tell face_recognition there is a missing face - something like facebook where one can click on the missing face? Or other mechanisms for pointing out a region of a missed face?

Thanks!

Mark

[ageitgey]

Instead of this:

face_locations = face_recognition.face_locations(image, number_of_times_to_upsample=2)

You could try the CNN model like this:

face_locations = face_recognition.face_locations(image, model="cnn")

It's more robust in difficult situations.

You can also pass in your own list of face_locations to face_encodings(), but that probably won't work very well unless the face boxes are perfect.

[pmi123]

Unfortunately, I don't have an nvidia graphics card to make that work. At least, I assume I need one to use the CNN model. I am running on a headless machine.

Mark

[ageitgey]

You don't need a gpu. It will just run more slowly without one.

[pmi123]

OK, I have CUDA and cudnn installed and dlib setup with them on Ubuntu 18.04. I tried face_locations = face_recognition.face_locations(image, model="cnn") and got this error:

File "/home/mark/.virtualenvs/memorabilia-JSON/lib/python3.6/site-packages/face_recognition/api.py", line 116, in face_locations
    return [_trim_css_to_bounds(_rect_to_css(face.rect), img.shape) for face in _raw_face_locations(img, number_of_times_to_upsample, "cnn")]
  File "/home/mark/.virtualenvs/memorabilia-JSON/lib/python3.6/site-packages/face_recognition/api.py", line 100, in _raw_face_locations
    return cnn_face_detector(img, number_of_times_to_upsample)
MemoryError: std::bad_alloc

I have this much RAM (in GB)

              total        used        free      shared  buff/cache   available
Mem:             15           7           4           0           2           6
Swap:             0           0           0

Any suggestions on how to fix this would be greatly appreciated.

Mark

[ageitgey]

try passing in face_locations = face_recognition.face_locations(image, model="cnn", number_of_times_to_upsample=0) and seeing if that fits in RAM.

[pmi123]

That worked, and tie software found the missing baby's face! number_of_times_to_upsample=0 and number_of_times_to_upsample=1 worked. number_of_times_to_upsample=2 failed. Hmmm 16 GB of total memory is too small. Time to upgrade! Thanks for all your help! Mark

[PLEX-GR00T]

Thank you so much guys, your conversation works for me CNN is not working in my lappy i have 8 GB ram, i tried the HOG and it works easily. i really am finding the reason and you and i got it, Thanks again.

[osean-man]

I'm using facerec_from_webcam_faster.py from the examples for some testing and if I change the model from hog to cnn, it doesn't recognize the face and lags hard. I'm also seeing my GPU usage happening but it's 10%-15% of the power. I'm not sure how to intelligently ask how to configure this to work with my GPU. I do have a Nvidia RTX3080. I have downloaded the CUDA drivers as well.