Stable Neighbor Denoising for Source-free Domain Adaptive Segmentation

We study source-free unsupervised domain adaptation (SFUDA) for semanticsegmentation, which aims to adapt a source-trained model to the target domainwithout accessing the source data. Many works have been proposed to addressthis challenging problem, among which uncertainty-based self-training is apredominant approach. However, without comprehensive denoising mechanisms, theystill largely fall into biased estimates when dealing with different domainsand confirmation bias. In this paper, we observe that pseudo-label noise ismainly contained in unstable samples in which the predictions of most pixelsundergo significant variations during self-training. Inspired by this, wepropose a novel mechanism to denoise unstable samples with stable ones.Specifically, we introduce the Stable Neighbor Denoising (SND) approach, whicheffectively discovers highly correlated stable and unstable samples by nearestneighbor retrieval and guides the reliable optimization of unstable samples bybi-level learning. Moreover, we compensate for the stable set by object-levelobject paste, which can further eliminate the bias caused by less learnedclasses. Our SND enjoys two advantages. First, SND does not require a specificsegmentor structure, endowing its universality. Second, SND simultaneouslyaddresses the issues of class, domain, and confirmation biases duringadaptation, ensuring its effectiveness. Extensive experiments show that SNDconsistently outperforms state-of-the-art methods in various SFUDA semanticsegmentation settings. In addition, SND can be easily integrated with otherapproaches, obtaining further improvements.