AttEntropy: On the Generalization Ability of Supervised Semantic Segmentation Transformers to New Objects in New Domains

TL;DR

This paper investigates how supervised vision transformers for semantic segmentation can generalize to unseen object categories and domains, revealing their ability to encode transferable information beyond training data.

Contribution

It demonstrates that supervised segmentation transformers inherently possess the ability to generalize to new objects and domains, a phenomenon previously observed only in unsupervised models.

Findings

Transformers trained for segmentation can segment unseen categories.

They generalize across diverse domains like lunar rocks and maritime hazards.

This ability is inherent and not solely due to training data.

Abstract

In addition to impressive performance, vision transformers have demonstrated remarkable abilities to encode information they were not trained to extract. For example, this information can be used to perform segmentation or single-view depth estimation even though the networks were only trained for image recognition. We show that a similar phenomenon occurs when explicitly training transformers for semantic segmentation in a supervised manner for a set of categories: Once trained, they provide valuable information even about categories absent from the training set. This information can be used to segment objects from these never-seen-before classes in domains as varied as road obstacles, aircraft parked at a terminal, lunar rocks, and maritime hazards.