linhuixiao/Awesome-Visual-Grounding

What it solves

This project provides a comprehensive, systematic survey of the field of visual grounding—the task of identifying specific objects in an image based on natural language descriptions. It addresses the need for a centralized, up-to-date resource that tracks a decade of research, including classical methods, new multimodal large language model (MLLM) approaches, and emerging research trends.

How it works

The repository serves as a curated database of visual grounding literature. It categorizes research into various settings (such as fully supervised, weakly supervised, and zero-shot), tracks representative works with links to their papers and code, and provides a clear comparison of different architectural approaches. It is designed to help both beginners and experienced researchers navigate the evolution of the field, from traditional CNN-based methods to modern Transformer and Vision-Language Pre-training (VLP) models.

Who it’s for

This resource is intended for computer vision researchers, students, and practitioners who want to understand the state of visual grounding, find benchmarks, or identify the latest advancements in referring expression comprehension and phrase grounding.

Highlights

• A comprehensive survey of visual grounding research spanning the last decade.
• Detailed categorization of methods including fully supervised, weakly supervised, semi-supervised, and zero-shot settings.
• Coverage of advanced topics like Grounding Multi-modal LLMs, generalized visual grounding, and VLP-based transfer learning.
• A curated list of representative works with direct links to papers and code implementations.
• Future research insights and a collection of relevant datasets and benchmarks.