Mitigating Hallucination in Visual Language Model Segmentation with Negative Sampling

Recent advancements in Multi Modal Language Models (MMLMs) have led to major breakthroughs in object reasoning segmentation, which plays an important role in human robot interaction. However, hallucination arises when these systems encounter the task of segmenting non-existing objects, for example, when asked to segment a cat which does not appear in an image, the system will output a patch of false positive pixels. In this paper, we utilizes negative sampling approach to address this issue. Unlike other approaches that require acquiring new negative training data, our approach does not require any extra data and tuning of only a few hundred steps is needed. We conduct extensive experiments to identify hallucination in MMLM based segmentation systems and mitigate false positive pixel rate in Language Instructed Segmentation Assistant (LISA) by 50% while maintaining similar performance in existing object segmentation.