Oscar: Object-Semantics Aligned Pre-training for Vision-and-Language Tasks VinVL: Revisiting Visual Representations in Vision-Language Models Updates