Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Revisiting Network Perturbation for Semi-Supervised Semantic Segmentation

Sien Li, Tao Wang, Ruizhe Hu, Wenxi Liu

TL;DR

By revisiting network perturbations, a new approach for network perturbation is introduced to expand the existing weak-to-strong consistency regularization for unlabeled data and a volatile learning process for labeled data is presented, which is uncommon in existing research.

Abstract

In semi-supervised semantic segmentation (SSS), weak-to-strong consistency regularization techniques are widely utilized in recent works, typically combined with input-level and feature-level perturbations. However, the integration between weak-to-strong consistency regularization and network perturbation has been relatively rare. We note several problems with existing network perturbations in SSS that may contribute to this phenomenon. By revisiting network perturbations, we introduce a new approach for network perturbation to expand the existing weak-to-strong consistency regularization for unlabeled data. Additionally, we present a volatile learning process for labeled data, which is uncommon in existing research. Building upon previous work that includes input-level and feature-level perturbations, we present MLPMatch (Multi-Level-Perturbation Match), an easy-to-implement and efficient framework for semi-supervised semantic segmentation. MLPMatch has been validated on the Pascal VOC and Cityscapes datasets, achieving state-of-the-art performance. Code is available from https://github.com/LlistenL/MLPMatch.

Revisiting Network Perturbation for Semi-Supervised Semantic Segmentation

TL;DR

By revisiting network perturbations, a new approach for network perturbation is introduced to expand the existing weak-to-strong consistency regularization for unlabeled data and a volatile learning process for labeled data is presented, which is uncommon in existing research.

Abstract

In semi-supervised semantic segmentation (SSS), weak-to-strong consistency regularization techniques are widely utilized in recent works, typically combined with input-level and feature-level perturbations. However, the integration between weak-to-strong consistency regularization and network perturbation has been relatively rare. We note several problems with existing network perturbations in SSS that may contribute to this phenomenon. By revisiting network perturbations, we introduce a new approach for network perturbation to expand the existing weak-to-strong consistency regularization for unlabeled data. Additionally, we present a volatile learning process for labeled data, which is uncommon in existing research. Building upon previous work that includes input-level and feature-level perturbations, we present MLPMatch (Multi-Level-Perturbation Match), an easy-to-implement and efficient framework for semi-supervised semantic segmentation. MLPMatch has been validated on the Pascal VOC and Cityscapes datasets, achieving state-of-the-art performance. Code is available from https://github.com/LlistenL/MLPMatch.

Paper Structure

This paper contains 14 sections, 6 equations, 6 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Semi-supervised Semantic Segmentation
  4. Network Perturbation
  5. Our Approach
  6. Preliminaries
  7. Weak-to-strong consistency based on network perturbation
  8. A volatile learning process for labeled data
  9. Our Holistic Framework: MLPMatch
  10. Experiments
  11. Experimental Setup
  12. Comparison with SOTAs
  13. Ablations studies
  14. Conclusion

Figures (6)

  • Figure 1: Performance comparison of MLPMatch with recent SOTA methods on the Pascal VOC dataset using a DeepLabv3plus and ResNet-101 model. The radii correspond to different quantities of labeled images.
  • Figure 2: An overview to MLPMatch. See comparisons with (a) FixMatch and (b) UniMatch in the training of unlabeled data on the left side. See comparisons with (a) FixMatch and (b) UniMatch in the training of labeled data on the right side. $A^w$ and $A^s$ represent the weak and strong input-level perturbations (augmentations).
  • Figure 3: Details of the BottleNeck in $\mathcal{\hat{F}}$. The convolutional forward pass of the BottleNeck is randomly activated or deactivated.
  • Figure 4: (a) Ablation - number of layers
  • Figure 5: (b) Ablation - $\lambda_{x}^{np}$
  • ...and 1 more figures