Native Segmentation Vision Transformers

TL;DR

This paper introduces Native Segmentation Vision Transformers, a novel architecture that uses content-aware spatial grouping layers to enable native, hierarchical segmentation within the backbone without additional segmentation heads, improving efficiency and zero-shot capabilities.

Contribution

The work presents a new backbone design with content-aware grouping layers that produce segmentation masks inherently, eliminating the need for extra segmentation-specific components.

Findings

Enables emergence of segmentation masks from grouping layers alone

Achieves strong zero-shot segmentation results without mask supervision

Offers a minimal, efficient model for downstream segmentation tasks

Abstract

Uniform downsampling remains the de facto standard for reducing spatial resolution in vision backbones. In this work, we propose an alternative design built around a content-aware spatial grouping layer, that dynamically assigns tokens to a reduced set based on image boundaries and their semantic content. Stacking our grouping layer across consecutive backbone stages results in hierarchical segmentation that arises natively in the feature extraction process, resulting in our coined Native Segmentation Vision Transformer. We show that a careful design of our architecture enables the emergence of strong segmentation masks solely from grouping layers, that is, without additional segmentation-specific heads. This sets the foundation for a new paradigm of native, backbone-level segmentation, which enables strong zero-shot results without mask supervision, as well as a minimal and efficient standalone model design for downstream segmentation tasks. Our project page is https://research.nvidia.com/labs/dvl/projects/native-segmentation.