Title: Domain Adaptation with a Single Vision-Language Embedding

Abstract

While domain adaptation in computer vision has seen extensive research, it typically necessitates access to target data during the training phase. This requirement poses significant challenges in real-world autonomous driving applications, particularly when dealing with rare events or adverse weather conditions. To address this limitation, we introduce a novel framework that relies on a single Vision-Language (VL) latent embedding rather than full target datasets.

Our approach centers on Prompt/Photo-driven Instance Normalization (PIN), a feature augmentation technique derived from contrastive language-image pre-training (CLIP) models. PIN optimizes the affine transformations of low-level source features to extract multiple visual styles from a single target VL latent embedding. This embedding can be generated in three ways: from a language prompt that describes the target domain, from a partially optimized language prompt, or from a single unlabeled image from the target domain.

We demonstrate that these mined styles serve as effective augmentations for both zero-shot (target-free) and one-shot unsupervised domain adaptation. Our experiments, conducted on semantic segmentation tasks using real-world driving datasets such as Cityscapes and ACDC (which features adverse conditions), validate the efficacy of our method. The results show that our approach surpasses relevant baselines in practical zero-shot and one-shot scenarios.

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