A Comparison on Visual Grounding

Visual grounding, also known as Referring Expression Comprehension (REC), is a fundamental multimodal AI task that involves localizing a specific object within an image based on a descriptive natural language query. This task is challenging as it requires models to parse complex, sometimes ambiguous, language and understand fine-grained visual details and relationships. The table below offers a detailed comparison of influential academic papers that have shaped this field. It summarizes each paper's core architectural schema—from early two-stage methods to modern Transformer-based and Large Vision-Language Models (LVLMs)—and delves into the specific techniques employed, such as Chain-of-Thought reasoning, Mixture-of-Experts, and self-consistent explanations. Furthermore, the table highlights the key contribution of each work and lists the primary benchmark datasets used for evaluation, including classics like RefCOCO and newer ones, providing a comprehensive overview of the field's evolution and key methodologies.