Enabling Multimodal Generation on CLIP via Vision-Language Knowledge Distillation

Benchmark Model Rank Results
image-captioning-on-coco-captionsVLKD (ViT-B/16)–BLEU-4: 16.7CIDER: 58.3METEOR: 19.7SPICE: 13.4
visual-question-answering-on-ok-vqaVLKD(ViT-B/16)–Accuracy: 10.5
visual-question-answering-on-vqa-v2-test-devVLKD–Accuracy: 44.5
visual-question-answering-on-vqa-v2-valVLKD(ViT-B/16)–Accuracy: 38.6