Unified Vision-Language Pre-Training for Image Captioning and VQA

24 Sep 2019  ·  Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason J. Corso, Jianfeng Gao ·

This paper presents a unified Vision-Language Pre-training (VLP) model. The model is unified in that (1) it can be fine-tuned for either vision-language generation (e.g., image captioning) or understanding (e.g., visual question answering) tasks, and (2) it uses a shared multi-layer transformer network for both encoding and decoding, which differs from many existing methods where the encoder and decoder are implemented using separate models. The unified VLP model is pre-trained on a large amount of image-text pairs using the unsupervised learning objectives of two tasks: bidirectional and sequence-to-sequence (seq2seq) masked vision-language prediction. The two tasks differ solely in what context the prediction conditions on. This is controlled by utilizing specific self-attention masks for the shared transformer network. To the best of our knowledge, VLP is the first reported model that achieves state-of-the-art results on both vision-language generation and understanding tasks, as disparate as image captioning and visual question answering, across three challenging benchmark datasets: COCO Captions, Flickr30k Captions, and VQA 2.0. The code and the pre-trained models are available at https://github.com/LuoweiZhou/VLP.

PDF Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Image Captioning COCO Captions test Unified VLP BLEU-4 36.5 # 2
CIDEr 116.9 # 1
METEOR 28.4 # 2
SPICE 21.2 # 1
Image Captioning Flickr30k Captions test Unified VLP BLEU-4 30.1 # 1
CIDEr 67.4 # 1
METEOR 23 # 1
SPICE 17 # 1
Visual Question Answering (VQA) VQA v2 test-std Unified VLP overall 70.7 # 27