Prompt-driven Multi-modal Unsupervised Domain Adaptation for 3D Semantic Segmentation

Existing multi-modal unsupervised domain adaptation (MM-UDA) methods focus on feature alignment to minimize distribution differences between source and target domains, but this can distort semantic structures and reduce visual feature discriminability. To solve this, we introduce PromptUDA, a prompt-driven MM-UDA method that leverages vision-language models to enhance visual feature discriminability using multi-modal prompts. PromptUDA comprises three crucial components: Multi-modal Data Preparation (MDP), Multi-modal Collaborative Interaction (MCI), and Cross-modal Cross-domain Prototype Contrastive Learning (CPCL). MDP employs a bidirectional fusion approach to process data, which facilitates better learning of multi-modal prompts. MCI enhances the interaction between domain-invariant prompt information and visual features, generating more discriminative semantic information. CPCL further explores the potential of visual features integrated with prompts, leveraging cross-domain and cross-modal advantages to learn domain-invariant features. Extensive experimental results demonstrate that our method outperforms state-of-the-art competitors in four domain adaptation scenarios.