Cross-domain 3D model classification via pseudo-labeling noise correction

Abstract

Unsupervised domain adaptation (UDA) with pseudo-labeling has become a key approach for cross-domain 3D model classification. Although it effectively narrows the gap between domains, the performance of existing UDA methods will drop significantly when applied to multi-category and multi-scene 3D model classification due to the dependence on 3D source domain labels and the impact of low-quality pseudo-labels. In this paper, we address this challenge by proposing an innovative cross-domain 3D model classification framework based on 2D–3D UDA and pseudo-label correction mechanism. Our method fully utilizes the rich semantic labels and scene information in the image domain for efficient image-to-3D cross-domain adaptation, completely eliminating the dependence on 3D labels. In addition, we introduce sufficient prior knowledge in the image domain to guide the adversarial training of the pseudo-label correction module. The introduction of cross-modal information improves the quality of pseudo-labels in cross-domain 3D classification, breaking the limitation of existing label denoising mechanisms that are limited to a single modality. Experimental results on multiple standard 3D model datasets and cross-domain generalization tasks show that this method outperforms existing mainstream 3D UDA methods in terms of robustness and classification performance, verifying its practicality and generalization ability without relying on 3D data annotation.