DVGBench: Implicit-to-Explicit Visual Grounding Benchmark in UAV Imagery with Large Vision-Language Models
🤖 AI Summary
Existing remote sensing visual grounding datasets predominantly rely on explicit referring expressions, which are ill-suited for implicit localization tasks requiring domain-specific knowledge. To address this gap, this work presents DVGBench—the first benchmark for implicit visual grounding in drone imagery—spanning six real-world application scenarios and introducing paired explicit–implicit queries. The authors propose DroneVG-R1, a model that integrates an Implicit-to-Explicit Chain-of-Thought (I2E-CoT) mechanism within a reinforcement learning framework to translate implicit references into actionable explicit expressions, thereby significantly enhancing localization performance. Experiments on DVGBench reveal that prevailing models struggle with implicit reasoning, whereas DroneVG-R1 achieves notably higher grounding accuracy.
Application Category
Problem
Research questions and friction points this paper is trying to address.
visual grounding
implicit referring expressions
UAV imagery
large vision-language models
remote sensing
Innovation
Methods, ideas, or system contributions that make the work stand out.
Implicit-to-Explicit Visual Grounding
Large Vision-Language Models
Chain-of-Thought Reasoning
UAV Imagery
Reinforcement Learning
🔎 Similar Papers
No similar papers found.
