accuracy with different face detector

yu4u / age-gender-estimation

Keras implementation of a CNN network for age and gender estimation

MIT License

1.46k stars 501 forks source link

accuracy with different face detector #65

Closed sbharadwajj closed 5 years ago

sbharadwajj commented 5 years ago

@yu4u Hi,

so i'm running the model with a different face detector(mtcnn) but made sure the img_size is 64 and the input to the model is of the form(batch_size, 64,64,3), but the age prediction is bad compared to using dlib.
The only difference with mtcnn is, its detecting more number of faces and the bounding boxes points vary. And I'm not resizing the image, but using its original dimensions. Can this affect the prediction of age?
What parameters affect the prediction of age/gender apart from the margin?

yu4u commented 5 years ago

Hi, @chichilicious

Is the face regions detected dlib and mtcnn are almost the same?
The inputs to the model are images with BGR color channels?
The margin around detected faces are correct (same)?

sbharadwajj commented 5 years ago

Hi,

yes, they are almost the same and I have performed these steps on them.

        for b bounding_boxes:
            b=[int(bi) for bi in b]
            x1,y1,x2,y2=b[0:4]
            x2 = x2 + 1
            y2 = y2 + 1
            w = y2 - y1
            h = x2 - x1
            xw1 = max(int(x1 - self.margin * w), 0)
            yw1 = max(int(y1 - self.margin * h), 0)
            xw2 = min(int(x2 + self.margin * w), img_w - 1)
            yw2 = min(int(y2 + self.margin * h), img_h - 1)

Yes, it is BGR
The margin is 0.4 only

yu4u commented 5 years ago

Hmm... One more thing, how about the range of pixel values (0-1 or 0-255).

If you save the inputs to the model created from both detectors, and the saved images are similar, there seems to be nothing wrong the code...

sbharadwajj commented 5 years ago

I verified that too. It is the same.

yu4u commented 5 years ago

😕

How "bad" are the results? If the results were worse than that of dlib but they are somewhat reasonable, it would be limitation of the model...

sbharadwajj commented 5 years ago

Actually, on further testing, I found out that they are somewhat reasonable compared to dlib. It lies in the same approximate range. But when I tested the model with another dataset, the performance was poor. But again that is probably the limitation of the model. Thanks for the help!