CVKD-UDA: Cross-View Knowledge Distillation for 3D Unsupervised Domain Adaptive Segmentation

2026-07-10

AI Summary

This work addresses the challenge of poor pseudo-label quality and weak generalization in unsupervised domain adaptation (UDA) for 3D segmentation, which arises from significant domain discrepancies. To this end, the authors propose a cross-view knowledge distillation framework centered on voxel size variation. By treating voxel scale as a core design element, the method constructs complementary viewpoint representations and incorporates a lightweight decoupled adapter along with an auxiliary imitative classifier to balance transferability and discriminability during knowledge transfer. Notably, this is the first approach to leverage voxel size variation explicitly for learning domain-invariant features. The proposed cross-view knowledge distillation mechanism substantially enhances the model’s perception capability in the target domain. Experiments demonstrate that the method significantly outperforms existing self-training approaches on two mainstream 3D UDA segmentation benchmarks, confirming its effectiveness and novelty.

Abstract

3D unsupervised domain adaptive (UDA) segmentation mitigates the high cost of manual annotations of the new domain data. Self-training has emerged as the dominant approach in this area, where its success heavily depends on a well-initialized warm-up model to generate reliable pseudo labels. However, existing methods often depend on source supervision or output-level adversarial alignment to obtain the warm-up model, which suffer from limited generalization and training instability due to the large domain gap between domains. Constructing domain-similar representations is an effective way to bridge this gap. In this work, we propose CVKD-UDA, which revisits voxel size as a core design factor to construct domain-similar representations and leverages cross-view complementary cues to balance transferability and discriminability of the warm-up model. First, we generate two complementary views by varying voxel sizes and introduce a cross-view knowledge distillation (CVKD) to enhance generalization and target perception of the model. Second, to balance transferability and discriminability, we design a lightweight Decouple-Adapter and an auxiliary imitation classifier to decouple cross-view knowledge transfer. Extensive experiments on two benchmarks demonstrate that CVKD-UDA effectively improves the performance of self-training methods and provides a new perspective for 3D UDA segmentation. Our code will be available at GitHub.

Problem

• 3D unsupervised domain adaptation
• domain gap
• warm-up model
• pseudo labeling
• self-training

Innovation

• Cross-View Knowledge Distillation
• Voxel Size
• Unsupervised Domain Adaptation
• 3D Segmentation
• Decouple-Adapter