Interesting, that makes sense. You still need a lot of diverse/realistic cropped objects that you can (realistically?) put into scenes. I'm not sure if simply dropping random objects into random scenes would work well enough.
Also, what exactly do the nets predict, since the foreground objects can have very complex contours, you'd need to predict an arbitrarily shaped polygon ?
Also, what exactly do the nets predict, since the foreground objects can have very complex contours, you'd need to predict an arbitrarily shaped polygon ?