shape_68.dat missing

[ctippur]

Nice work. I was trying to run the sample code. I seem to be missing shape_68.dat file. I see that you have tried to add it but I dont see the file.

[vardanagarwal]

Thanks @ctippur. I switched from using Dlib facial landmark model to a TensorFlow model which I am pretty sure you can find in the repo. However, if you want to continue to use Dlib you can download the model from here: https://github.com/davisking/dlib-models/blob/master/shape_predictor_68_face_landmarks.dat.bz2.

Hope that solves the issue.

[ctippur]

Thanks @vardanagarwal. I am looking for the most accurate method to determine the location points. If I understand right, TF model is a better way to go?

[vardanagarwal]

Yeah, based on my testing I felt is was better. You can have a look at the results here: https://towardsdatascience.com/robust-facial-landmarks-for-occluded-angled-faces-925e465cbf2e?sk=505eb1101576227f4c38474092dd4c22

[ctippur]

Vardan,

Thank you. Thats quite impressive. I modified contouring function to see if I can get just the center of pupil coordinates.

I seem to be getting a lot of (<class 'ValueError'>, ValueError('max() arg is an empty sequence'), <traceback object at 0x148fe4b90>) I would also like to improve the precision. Any ideas on how I can improve the precision and not lose frames?

try:
        cnt = max(cnts, key = cv2.contourArea)
        M = cv2.moments(cnt)
        cx = int(M['m10']/M['m00'])
        cy = int(M['m01']/M['m00'])

        if right:
            cx += mid

        return eye, cx, cy
    except:
        print (sys.exc_info())
        return eye, None, None

[vardanagarwal]

That's to do with the threshold. If your threshold is very small then no contour will be detected leading to that error. You can add an if condition checking that length of cnts>0.

I am trying to figure out to automate the thresholding part then this won't be an issue.

Can you elaborate what you mean by precision here and why you feel it might lose frames if it is tried to be improved.

[ctippur]

You are right. The threshold for the most part seem to be 0. I will have to look at the precision question again. I have a still video and I was hoping to see the cx and cy to pretty much be the same. Since so many frames are being rejected.

To increase the threshold, I am looking at some variatiobs below. Let me know if we can collaborate.

    thresh = cv2.erode(thresh, None, iterations=2) 
    thresh = cv2.dilate(thresh, None, iterations=4) 
    thresh = cv2.medianBlur(thresh, 3) 
    thresh = cv2.bitwise_not(thresh)

[vardanagarwal]

Yeah sure! As mentioned earlier it is currently controlled by a trackbar which means it requires manual work so I am definitely up for it.

[ctippur]

Quick benchmarking

| Threshold   |      Left x      |  left y | right x | right y |
|----------|:-------------:|------:|------:|------:|
| 1 |  180 | 180| 199 | 199 |
|3|165|165|180|180|

[vardanagarwal]

It would be better to close this issue and open up another one having the apt name so that anyone else looking to contribute can find it.

[ctippur]

Perfect. Closing this to reopen another issue.