CLIP has seen zero segmentations or bounding-boxed labeled examples of any object classes. Pure image-level contrastive learning against (extremely) noisy natural-language text captions.