Colorization is far from perfect indeed. I'm not sure how they model the temporal dimension (recognizing the same object in the prev/next frame and painting it

Colorization is far from perfect indeed. I'm not sure how they model the temporal dimension (recognizing the same object in the prev/next frame and painting it of the same color - unless it has to be different because of shadows, lighting change, etc.) in these neural nets, does anyone know?