How to train the DAN-Tracking.npz model？

[zhaofang216]

Hi ,

  Thanks for your codes.

  I have trained a model of 106  landmarks (without tracking) via your codes,  now I want to train the tracking model, but i do not know how to do it.  Can you give me some suggestions?

[MarekKowalski]

Hi,

I assume, that by tracking model you mean the one that is able to predict whether it lost tracking or not. This model was trained by adding a single additional dense layer attached to the penultimate layer of the first stage, you can see it in line 117 of this file.

The layer was trained to distinguish faces from non-faces using the WIDER FACE dataset. Since that dataset contains annotations for all faces in all images, it was easy to generate a large dataset of face and non-face patches. In addition we blurred some of the face patches to simulate motion blur that occurs during tracking.

Does that help?

Marek

[zhaofang216]

Hi Marek, Thanks for you kindly reply, and could you give me more details about how to train additional layer? I wonder what the neural network to train the confidence layer is. How many layers the neural network contains?

Thanks.

[MarekKowalski]

Hi, As mentioned in the previous post, the neural network consists of a single layer, you can see it in the file I linked to.

Marek