Using dlib for detection about face_classification

[oarriaga]

I believe the detection is working properly, and dlib will probably not increase the accuracy of the classifications. However, it might be better for detection when it comes to slight head rotations which are problematic in the current openCV face detection function.

[Zumbalamambo]

why is that it is not detecting asian faces? is it because imdb has only westerners faces?

[oarriaga]

I didn't know this problem existed. If is not detecting correctly any sort of faces it only depends on the openCV detection module. It might work better with dlib. I will gladly accept a pull request with this change.

[oarriaga]

BTW, when you say detection is not working you are referring to the fact that bounding boxes are not being drawn in Asian faces? Or that the bounding boxes are there but the classification of the emotion/gender is not working?

[Zumbalamambo]

it can draw bounding boxes. However,recognition is not accurate. I have tried imwrite the cropped face that Im passing to the model. It is not recognizing any of such cropped asian faces accurately

[Zumbalamambo]

If I have to add images to the imdb with asian faces, will it work? if so How do I go about sorting the dataset so I can use it for training

[Zumbalamambo]

imdb has very very less amount of asian faces

[oarriaga]

Yes if you expand your data it will help.

[Zumbalamambo]

Alright I will openup pull request expanded model. Do you have any idea about how data has to be organized as in imdb so I can retrain the existing model?

[oarriaga]

Please look at the source code I believe it shouldn't be to complicated to understand. The inputs shape is organized as follow (num_samples, height, width, num_channels) the outputs shape (num_samples, num_classes).

[Zumbalamambo]

It makes use of imdb.mat file. How do I annotate the dataset with mat file?