ERNIE-ViL 2.0: Multi-view Contrastive Learning for Image-Text Pre-training

30 Sep 2022  ยท  Bin Shan, Weichong Yin, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang ยท

Recent Vision-Language Pre-trained (VLP) models based on dual encoder have attracted extensive attention from academia and industry due to their superior performance on various cross-modal tasks and high computational efficiency. They attempt to learn cross-modal representation using contrastive learning on image-text pairs, however, the built inter-modal correlations only rely on a single view for each modality. Actually, an image or a text contains various potential views, just as humans could capture a real-world scene via diverse descriptions or photos. In this paper, we propose ERNIE-ViL 2.0, a Multi-View Contrastive learning framework to build intra-modal and inter-modal correlations between diverse views simultaneously, aiming at learning a more robust cross-modal representation. Specifically, we construct multiple views within each modality to learn the intra-modal correlation for enhancing the single-modal representation. Besides the inherent visual/textual views, we construct sequences of object tags as a special textual view to narrow the cross-modal semantic gap on noisy image-text pairs. Pre-trained with 29M publicly available datasets, ERNIE-ViL 2.0 achieves competitive results on English cross-modal retrieval. Additionally, to generalize our method to Chinese cross-modal tasks, we train ERNIE-ViL 2.0 through scaling up the pre-training datasets to 1.5B Chinese image-text pairs, resulting in significant improvements compared to previous SOTA results on Chinese cross-modal retrieval. We release our pre-trained models in https://github.com/PaddlePaddle/ERNIE.

PDF Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data
Result Benchmark
Image Retrieval AIC-ICC ERNIE-ViL2.0 Recall@1 19.0 # 1
Recall@10 43.5 # 1
Recall@5 35.3 # 2
Image-to-Text Retrieval AIC-ICC ERNIE-ViL2.0 Recall@1 33.7 # 1
Recall@5 52.1 # 1
Recall@10 60.0 # 1
Cross-Modal Retrieval COCO 2014 ERNIE-ViL 2.0 Image-to-text R@1 77.4 # 13
Image-to-text R@10 97.1 # 11
Image-to-text R@5 93.6 # 13
Text-to-image R@1 59.5 # 17
Text-to-image R@10 90.1 # 13
Text-to-image R@5 83.4 # 15
Zero-Shot Cross-Modal Retrieval COCO 2014 ERNIE-ViL 2.0 Image-to-text R@1 63.1 # 11
Image-to-text R@5 85.7 # 11
Image-to-text R@10 91.4 # 10
Text-to-image R@1 46.0 # 11
Text-to-image R@5 71.4 # 10
Text-to-image R@10 80.4 # 11
Zero-shot Text-to-Image Retrieval COCO-CN ERNIE-ViL 2.0 Recall@1 69.6 # 2
Recall@5 91.2 # 2
Recall@10 96.9 # 2
Zero-shot Image Retrieval COCO-CN ERNIE-ViL 2.0 R@1 69.6 # 3
R@5 91.2 # 4
R@10 96.9 # 3
Cross-Modal Retrieval Flickr30k ERNIE-ViL 2.0 Image-to-text R@1 97.2 # 5
Image-to-text R@10 100.0 # 1
Image-to-text R@5 100.0 # 1
Text-to-image R@1 93.3 # 1
Text-to-image R@10 99.8 # 1
Text-to-image R@5 99.4 # 1
Image-to-Text Retrieval Flickr30k ERNIE-ViL 2.0 Recall@1 96.1 # 6
Recall@5 99.9 # 6
Recall@10 100.0 # 1
Zero-Shot Cross-Modal Retrieval Flickr30k ERNIE-ViL 2.0 Image-to-text R@1 91.2 # 6
Image-to-text R@5 99.1 # 9
Image-to-text R@10 99.8 # 5
Text-to-image R@1 77.4 # 9
Text-to-image R@5 93.8 # 10
Text-to-image R@10 96.4 # 11

Methods