Title: Unsupervised Collaborative Domain Adaptation for Driving Scene Parsing

Abstract

The ability to accurately parse driving scenes is a cornerstone of autonomous vehicle functionality, particularly within open and evolving operational environments. Nevertheless, adapting perception systems to novel deployment domains presents significant hurdles. Obtaining pixel-level annotations is costly, and source-domain data is frequently unavailable due to privacy regulations, security protocols, or ownership restrictions. Current source-free unsupervised domain adaptation techniques generally depend on a solitary pre-trained source model. This reliance renders the resulting perception system susceptible to biases inherent to the source domain, thereby compromising robustness across varying road layouts, lighting, weather, and traffic scenarios.

To address these limitations, this study introduces an Unsupervised Collaborative Domain Adaptation (UCDA) framework designed for source-free driving scene parsing. This approach facilitates the transfer of complementary knowledge from several pre-trained source models into a single unified target model, all without requiring access to original source samples. To harmonize predictions generated by independently trained models, UCDA employs a class-level prototype memory bank. It gauges cross-model prediction reliability by measuring prototype similarity, which mitigates discrepancies in confidence scales among different source models.

Leveraging this complementary supervision, UCDA implements a two-stage transfer strategy. Initially, multiple source models undergo refinement on unlabeled target-domain driving data via collaborative optimization, utilizing both positive and negative consistency constraints. Subsequently, the validated expertise of these models is distilled into a single, deployable target model. Extensive evaluations conducted on public driving-scene datasets and real-world data gathered from an autonomous vehicle platform indicate that UCDA successfully integrates complementary multi-source knowledge. This integration enhances the reliability and generalization of target-domain scene parsing across a wide array of driving environments.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC