predicted_genders[i][0] < 0.5 or > 0.5 for male on video

[N. YC. Kim]

Hi Yu-san,

Thanks for this repo. I have a question regarding this snippet in demo.py
"M" if predicted_genders[i][0] < 0.5 else "F")
I am predicting on one face only so my code is:

face[0, :, :, :] = cv2.resize(img[yw1:yw2 + 1, xw1:xw2 + 1, :], (imgsize, imgsize))
crop = img[y1: y2+1, x1:x2+1]
height, width, _ = crop.shape
result = self.model.predict(face)
predicted_gender = result[0]
print("gender {}".format(predicted_gender[0][0]))
if predicted_gender[0][0] < 0.5:  #is this correct?
     pixelated = pixelate(crop, width, height)

I'm getting this output showing all female:
gender 0.7589111924171448
gender 0.6748817563056946
gender 0.7280391454696655
gender 0.7113814353942871
gender 0.6321226954460144
gender 0.6755571961402893
gender 0.655595600605011
gender 0.6332910060882568
gender 0.665791392326355
gender 0.6619336009025574
frame 10, 2 percent done
gender 0.6297609210014343
gender 0.6310742497444153
gender 0.6388582587242126
gender 0.6477266550064087
gender 0.6689814329147339
gender 0.6472148299217224
gender 0.6857687830924988
gender 0.6730141639709473
gender 0.675566554069519
gender 0.6839122772216797
gender 0.6692551374435425
gender 0.6698795557022095
gender 0.6731051802635193
gender 0.6431819796562195
gender 0.7187308073043823
gender 0.7388319373130798

For video of myself here face detected with mtcnn which bbox I think is similar to dlib.
https://youtu.be/QMs9DZk0IOs

With the sign reversed, ie. if predicted_gender[0][0] > 0.5, gender prediction is almost perfect even on profile: https://www.youtube.com/watch?v=4yxAbq-3avk&feature=youtu.be

So I don't understand how I'm being detected as female because when I ran your demo.py on webcam on myself, it detected me as male 100% but it was not good on profile face due to dlib limitations.

My function is below but I think the above snippet is enough:

    def predict_gender(self, img, bboxes):
        #check img
        #print(img.shape)
        #pad with margin around image before gender detection
        #params from argparse
        #margin = 0.4 #for imdb
        margin = 0.4
        imgsize = 64 #default 64
        
        img_h, img_w, _ = np.shape(img)
        #detect faces with DSFD and resize to 64 by 64, but resize to 128 (retrain at this size)
        #with VGGFace img size is 224*224
        face = np.empty((1, imgsize, imgsize, 3))
        
        if len(bboxes) > 0:
            for i in range(len(bboxes)):
                x1, y1 = bboxes[i][0], bboxes[i][1]
                w, h = bboxes[i][2], bboxes[i][3]
                x2, y2 = x1 + w, y1 + h 
                xw1 = max(int(x1 - margin * w), 0)
                yw1 = max(int(y1 - margin * h), 0)
                xw2 = min(int(x2 + margin * w), img_w - 1)
                yw2 = min(int(y2 + margin * h), img_h - 1)
                
                face[0, :, :, :] = cv2.resize(img[yw1:yw2 + 1, xw1:xw2 + 1, :], (imgsize, imgsize))
                crop = img[y1: y2+1, x1:x2+1]
                height, width, _ = crop.shape
                result = self.model.predict(face)
                predicted_gender = result[0]
                print("gender {}".format(predicted_gender[0][0]))
                if predicted_gender[0][0] < 0.5: #reversed to >0.5 for second Youtube video result
                    pixelated = pixelate(crop, width, height)
                    img[y1:y2+1, x1:x2+1] = pixelated
        
        return img

[Yusuke Uchida]

What does happen if you simply run demo.py? The same results?

[N. YC. Kim]

No, when I ran demo.py, it was predicting correctly. I did not have to reverse the > sign. But I'll try again.

[gggmt]

Hello, Have you evaluated the gender on wiki datasets ?
I was very confused because I didn't get good results.