Large-scale vision-language pre-trained (VLP) models are prone to hallucinate non-existent visual objects when generating text based on visual information. In this paper, we exhaustively probe the object hallucination problem from three aspects. First, we examine various state-of-the-art VLP models, showing that models achieving better scores on standard metrics(e.g., BLEU-4, CIDEr) could hallucinate objects more frequently. Second, we investigate how different types of visual features in VLP influence hallucination, including region-based, grid-based, and patch-based. Surprisingly, we find that patch-based features perform the best and smaller patch resolution yields a non-trivial reduction in object hallucination. Third, we decouple various VLP objectives and demonstrate their effectiveness in alleviating object hallucination. Based on that, we propose a new pre-training loss, object masked language modeling, to further reduce object hallucination. We evaluate models on both COCO (in-domain) and NoCaps (out-of-domain) datasets with our improved CHAIR metric. Furthermore, we investigate the effects of various text decoding strategies and image augmentation methods on object hallucination.

该论文系统研究了视觉语言预训练模型中对象幻觉问题，从近期最先进的VLP模型，VLP中不同类型的图像编码方式，以及VLP目标的不同方面入手，提出了一种名为ObjMLM的简单而有效的VLP损失，能够减少对象幻觉。实验表明，ObjMLM可以将对象幻觉降低多达17.4%。

可信并非忠实：探究视觉-语言预训练中对象幻觉问题