Acuracy on Gender model

[Huckaby__Tim]

Hola, we are running the current version of OV.  we have been using the gender model for over a year pretty successfully.  i might be insane, but it just seems like it's accuracy has degraded since we first started using it.  we have built a number of test tools to strip everything down except for simply the gender model (plus the face detector) and it just is not doing well in terms of accuracy.  

so, questions:

- do you intel actually improve on the "canned" ov models like gender or are they static?

- we are using a myriad of different face detectors in different use cases.  might it be that the gender model needs a very specific crop of the face or resolution for it' to perform with accuracy?

- any suggestions of another gender model that may have better accuracy?

thanks in advance.  -tim

[Shubha_R_Intel]

Dear Huckaby, Tim,

You are definitely not insane.

Please look at the intel model downloads site here . The last modified date will tell you whether the model has been modified or not.

Also, pre-processing images before training the model - but failing to tell Model Optimizer about it -  is a common reason behind why customers experience accuracy loss. Yes if you crop your images before training then you had better tell model optimizer about it through --input_shape. 

Please see my detailed answer to this post

Thanks !

Shubha