Few-shot Class-Incremental Semantic Segmentation via Pseudo-Labeling and Knowledge Distillation

TL;DR

This paper proposes a novel method for few-shot class-incremental semantic segmentation that uses pseudo-labeling and knowledge distillation to effectively learn new classes from limited data while retaining existing knowledge.

Contribution

It introduces a coarse-to-fine pseudo-labeling strategy combined with knowledge distillation within a unified network for improved few-shot segmentation.

Findings

Effective pseudo-labeling improves novel class learning.

Knowledge distillation helps prevent catastrophic forgetting.

Method outperforms baselines on Cityscapes and KITTI datasets.

Abstract

We address the problem of learning new classes for semantic segmentation models from few examples, which is challenging because of the following two reasons. Firstly, it is difficult to learn from limited novel data to capture the underlying class distribution. Secondly, it is challenging to retain knowledge for existing classes and to avoid catastrophic forgetting. For learning from limited data, we propose a pseudo-labeling strategy to augment the few-shot training annotations in order to learn novel classes more effectively. Given only one or a few images labeled with the novel classes and a much larger set of unlabeled images, we transfer the knowledge from labeled images to unlabeled images with a coarse-to-fine pseudo-labeling approach in two steps. Specifically, we first match each labeled image to its nearest neighbors in the unlabeled image set at the scene level, in order to obtain images with a similar scene layout. This is followed by obtaining pseudo-labels within this neighborhood by applying classifiers learned on the few-shot annotations. In addition, we use knowledge distillation on both labeled and unlabeled data to retain knowledge on existing classes. We integrate the above steps into a single convolutional neural network with a unified learning objective. Extensive experiments on the Cityscapes and KITTI datasets validate the efficacy of the proposed approach in the self-driving domain. Code is available from https://github.com/ChasonJiang/FSCILSS.