Oscar: Object-Semantics Aligned Pre-training for Vision-Language Tasks - View it on GitHub