VLT: Vision-Language Transformer and Query Generation for Referring Segmentation

TL;DR

This paper introduces VLT, a vision-language transformer with dynamic query generation and balancing modules, improving referring segmentation by better understanding diverse language expressions and interactions with images.

Contribution

The paper presents a novel dynamic query generation and balancing framework for referring segmentation, enabling better handling of language diversity and multi-modal interactions.

Findings

Achieves state-of-the-art results on five datasets.

Effectively models diverse language expressions.

Improves understanding of vision-language interactions.

Abstract

We propose a Vision-Language Transformer (VLT) framework for referring segmentation to facilitate deep interactions among multi-modal information and enhance the holistic understanding to vision-language features. There are different ways to understand the dynamic emphasis of a language expression, especially when interacting with the image. However, the learned queries in existing transformer works are fixed after training, which cannot cope with the randomness and huge diversity of the language expressions. To address this issue, we propose a Query Generation Module, which dynamically produces multiple sets of input-specific queries to represent the diverse comprehensions of language expression. To find the best among these diverse comprehensions, so as to generate a better mask, we propose a Query Balance Module to selectively fuse the corresponding responses of the set of queries. Furthermore, to enhance the model's ability in dealing with diverse language expressions, we consider inter-sample learning to explicitly endow the model with knowledge of understanding different language expressions to the same object. We introduce masked contrastive learning to narrow down the features of different expressions for the same target object while distinguishing the features of different objects. The proposed approach is lightweight and achieves new state-of-the-art referring segmentation results consistently on five datasets.