Improving Few-Shot Segmentation Using Negative Class Instances

Keywords

Loading...
Thumbnail Image

Issue Date

2023-07-25

Language

en

Document type

Journal Title

Journal ISSN

Volume Title

Publisher

Title

ISSN

Volume

Issue

Startpage

Endpage

DOI

Abstract

Existing Few-Shot Semantic Segmentation (FSS) networks only rely on training with query images containing the class of interest, which hinders their performance in scenarios where the class is absent, compromising the network’s ecological validity. To rectify this, we propose a novel training loop that incorporates negative instances during training, providing the networks with exposure to a broader range of class instances. After testing this approach, our results show that not only did the models trained on our novel training loop improve in cases where the class of interest was absent from the query image, the models also improved when the class of interest was present.

Description

Citation

Faculty

Faculteit der Sociale Wetenschappen