Smoothing Matters: Momentum Transformer for Domain Adaptive Semantic Segmentation

TL;DR

This paper introduces a momentum-based smoothing technique and dynamic discrepancy measurement to improve domain adaptive semantic segmentation with Vision Transformers, addressing high-frequency noise issues and enhancing transferability.

Contribution

It proposes a novel momentum network and dynamic discrepancy measurement to enhance the transferability of local ViTs in domain adaptive semantic segmentation.

Findings

Outperforms state-of-the-art methods on sim2real benchmarks

Effectively reduces high-frequency noise in features and pseudo labels

Improves the transferability of local Vision Transformers

Abstract

After the great success of Vision Transformer variants (ViTs) in computer vision, it has also demonstrated great potential in domain adaptive semantic segmentation. Unfortunately, straightforwardly applying local ViTs in domain adaptive semantic segmentation does not bring in expected improvement. We find that the pitfall of local ViTs is due to the severe high-frequency components generated during both the pseudo-label construction and features alignment for target domains. These high-frequency components make the training of local ViTs very unsmooth and hurt their transferability. In this paper, we introduce a low-pass filtering mechanism, momentum network, to smooth the learning dynamics of target domain features and pseudo labels. Furthermore, we propose a dynamic of discrepancy measurement to align the distributions in the source and target domains via dynamic weights to evaluate the importance of the samples. After tackling the above issues, extensive experiments on sim2real benchmarks show that the proposed method outperforms the state-of-the-art methods. Our codes are available at https://github.com/alpc91/TransDA