This paper presents a comprehensive survey of recent general-purpose vision language models (VLMs) focusing on visual grounding. Visual grounding refers to the model's ability to identify regions within visual input that match textual descriptions. It has applications in a wide range of fields, including understanding referential representations, answering questions about details in images or videos, and low- and high-level control in simulations and real-world environments. This paper outlines the importance of grounding in VLMs, describes the core components of a modern paradigm for grounding model development, and examines real-world applications, including benchmarks and evaluation metrics for grounding multimodal models. Furthermore, we discuss the multifaceted interrelationships between visual grounding, multimodal thought chains, and the inference of VLMs. We analyze the unique challenges of visual grounding and suggest promising directions for future research.
