PEVL: Position-enhanced Pre-training and Prompt Tuning for Vision-language Models

23 May 2022  ยท  Yuan YAO, Qianyu Chen, Ao Zhang, Wei Ji, Zhiyuan Liu, Tat-Seng Chua, Maosong Sun ยท

Vision-language pre-training (VLP) has shown impressive performance on a wide range of cross-modal tasks, where VLP models without reliance on object detectors are becoming the mainstream due to their superior computation efficiency and competitive performance. However, the removal of object detectors also deprives the capability of VLP models in explicit object modeling, which is essential to various position-sensitive vision-language (VL) tasks, such as referring expression comprehension and visual commonsense reasoning. To address the challenge, we introduce PEVL that enhances the pre-training and prompt tuning of VLP models with explicit object position modeling. Specifically, PEVL reformulates discretized object positions and language in a unified language modeling framework, which facilitates explicit VL alignment during pre-training, and also enables flexible prompt tuning for various downstream tasks. We show that PEVL enables state-of-the-art performance of detector-free VLP models on position-sensitive tasks such as referring expression comprehension and phrase grounding, and also improves the performance on position-insensitive tasks with grounded inputs. We make the data and code for this paper publicly available at https://github.com/thunlp/PEVL.

PDF Abstract
Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data
Result Benchmark
Phrase Grounding Flickr30k Entities Dev PEVL R@1 84.1 # 2
Phrase Grounding Flickr30k Entities Test PEVL R@1 84.4 # 4
Visual Question Answering (VQA) GQA PEVL+ Accuracy 77 # 1
Visual Commonsense Reasoning VCR (Q-A) dev PEVL Accuracy 75.1 # 1
Visual Commonsense Reasoning VCR (Q-AR) dev PEVL Accuracy 57.8 # 1
Visual Commonsense Reasoning VCR (QA-R) dev PEVL Accuracy 76.4 # 1
Visual Commonsense Reasoning VCR (Q-AR) test PEVL Accuracy 58.6 # 1
Visual Commonsense Reasoning VCR (QA-R) test PEVL Accuracy 76.7 # 1
Visual Commonsense Reasoning VCR (Q-A) test PEVL Accuracy 76.0 # 1
Visual Relationship Detection Visual Genome PEVL R@50 64.4 # 1
R@100 66.3 # 1
mR@50 21.7 # 1
mR@100 23.5 # 1

Methods


No methods listed for this paper. Add relevant methods here