Can interchanging of image sets A & B between training help in better prediction?

[ugramund]

I was training with around 300+ images for 2 different persons on sides A & B for some time but noticed that some faces of side B were not at all predicted well on side A in the preview window because those facial angles & expressions were missing in my B image set.

Then I interchanged my sets i.e the image set used for A was then used for B & vice versa. After some hours of training I saw those faces got predicted very well & so I again interchanged my image sets back to their original A & B location. This time, after training for an hour or so, those set B faces got perfectly predicted on side A.

The question I want to ask is, did interchanging of image sets A & B helped me in better prediction or was it because of simply more training?

An extra question I would like to add is,

If my set B images lacks some facial angles & expressions compared to set A images, then if I train for some time with an image set of a different person on side B but with all the required facial angles & expressions of set A images & then again revert back to my original set B images, will this help in any better prediction of images & output of my training?

[torzdf]

It can help, it can make things worse. Ultimately your mileage may vary...

As you've seen it can help "fill gaps" where there are not matching poses on the other side. The downside of switching data is that identity can bleed, so that your swap ends up looking less like the target and more like a mix of the 2 faces.

[bryanlyon]

(Adding on to Torzdf's very good response)

I think it's important to note that even if you don't have full coverage of all poses, as long as you have enough variety it should be able to fill the gap given enough training.