Referring Expressions as a Lens into Spatial Language Grounding in Vision-Language Models

TL;DR

This paper evaluates vision-language models' spatial reasoning abilities using Referring Expression Comprehension, revealing their strengths and weaknesses in understanding complex, ambiguous, and negated spatial expressions.

Contribution

It introduces the Referring Expression Comprehension task as a new platform for analyzing spatial grounding in vision-language models, focusing on complex and ambiguous spatial language.

Findings

Models struggle with complex spatial expressions.

Performance varies across spatial categories.

Challenges are influenced by model architecture and expression complexity.

Abstract

Spatial Reasoning is an important component of human cognition and is an area in which the latest Vision-language models (VLMs) show signs of difficulty. The current analysis works use image captioning tasks and visual question answering. In this work, we propose using the Referring Expression Comprehension task instead as a platform for the evaluation of spatial reasoning by VLMs. This platform provides the opportunity for a deeper analysis of spatial comprehension and grounding abilities when there is 1) ambiguity in object detection, 2) complex spatial expressions with a longer sentence structure and multiple spatial relations, and 3) expressions with negation ('not'). In our analysis, we use task-specific architectures as well as large VLMs and highlight their strengths and weaknesses in dealing with these specific situations. While all these models face challenges with the task at hand, the relative behaviors depend on the underlying models and the specific categories of spatial semantics (topological, directional, proximal, etc.). Our results highlight these challenges and behaviors and provide insight into research gaps and future directions.