Prismer: A Vision-Language Model with Multi-Task Experts

Benchmark Model Rank Results
image-captioning-on-nocaps-valPrismer#1CIDEr: 107.9SPICE: 14.8
image-captioning-on-nocaps-entirePrismer#2CIDEr: 110.84B1: 84.87B2: 69.99B3: 52.48B4: 33.66…
visual-question-answering-on-vqa-v2-test-stdPrismer#12overall: 78.49yes/no: 93.09number: 61.39other: 69.70
visual-question-answering-on-vqa-v2-test-devPrismer#15Accuracy: 78.43
image-captioning-on-coco-captionsPrismer#17BLEU-4: 40.4CIDER: 136.5METEOR: 31.4SPICE: 24.4