Evaluating Object Hallucination in Large Vision-Language Models

Explore this paper's citation graph

Summary

This work presents the first systematic study on object hallucination of LVLMs, and designs an improved evaluation method by proposing a polling-based query method called POPE, which can evaluate the object hallucinated objects in a more stable and flexible way.

Type
preprint
Published
2023-05-17
Cited by
1,922
References
54
Access
Open access

Keywords

Hallucinating, Object (grammar), Computer science, Visual Hallucination, Polling

References

Cited by

Related papers