Regression Example in Documentation

In the documentation related to the BIWI Head Pose Regression, the objective of the regression is stated as:

"On pictures of persons, we have to find the center of their face."
However, in the example prediction provided, the red dot is shown way below the image (see screenshot below).

Is there a reason for the prediction to vary so widely from the ground truth?

Do note the documentation models are run very briefly, so they won’t be as accurate :slight_smile: I have a feeling this may be the case here.

Thanks … that’s probably what’s happening. I was just curious because the prediction appears to be a really long way off from the ground truth … even going so far as outside the image.

When I tried using the model trained during the class (Week 3, BIWI Head Pose) on other face photos I got similarly inaccurate results inspite of training for 5 epochs (learn.fit_one_cycle(5, slice(lr)). An example result I got is below:

You need to think about how the image looks compared to what was being trained. To me, this image is very different because there’s a different subject (gender bias) along with a completely different background. So data context is important with problems like this I’ve found. :slight_smile:

1 Like