Synth-to-real semi-supervised learning for visual tasks