Normalized Feature Distillation for Semantic Segmentation

TL;DR

This paper introduces a normalized feature distillation method for semantic segmentation that simplifies knowledge transfer by focusing on feature normalization, leading to state-of-the-art results without complex manual design.

Contribution

Proposes a simple normalized feature distillation approach that improves semantic segmentation performance without elaborate manual knowledge design.

Findings

Achieves state-of-the-art results on Cityscapes, VOC 2012, and ADE20K datasets.

Simplifies knowledge distillation by normalization, removing the need for manual feature design.

Effective for model compression in semantic segmentation.

Abstract

As a promising approach in model compression, knowledge distillation improves the performance of a compact model by transferring the knowledge from a cumbersome one. The kind of knowledge used to guide the training of the student is important. Previous distillation methods in semantic segmentation strive to extract various forms of knowledge from the features, which involve elaborate manual design relying on prior information and have limited performance gains. In this paper, we propose a simple yet effective feature distillation method called normalized feature distillation (NFD), aiming to enable effective distillation with the original features without the need to manually design new forms of knowledge. The key idea is to prevent the student from focusing on imitating the magnitude of the teacher's feature response by normalization. Our method achieves state-of-the-art distillation results for semantic segmentation on Cityscapes, VOC 2012, and ADE20K datasets. Code will be available.