Abstract:Unsupervised Domain Adaptation (UDA) is crucial for reducing the need for extensive manual data annotation when training deep networks on point cloud data. A significant challenge of UDA lies in effectively bridging the domain gap. To tackle this challenge, we propose \textbf{C}urvature \textbf{D}iversity-Driven \textbf{N}uclear-Norm Wasserstein \textbf{D}omain Alignment (CDND). Our approach first introduces a \textit{\textbf{Curv}ature Diversity-driven Deformation \textbf{Rec}onstruction (CurvRec)} task, which effectively mitigates the gap between the source and target domains by enabling the model to extract salient features from semantically rich regions of a given point cloud. We then propose \textit{\textbf{D}eformation-based \textbf{N}uclear-norm \textbf{W}asserstein \textbf{D}iscrepancy (D-NWD)}, which applies the Nuclear-norm Wasserstein Discrepancy to both \textit{deformed and original} data samples to align the source and target domains. Furthermore, we contribute a theoretical justification for the effectiveness of D-NWD in distribution alignment and demonstrate that it is \textit{generic} enough to be applied to \textbf{any} deformations. To validate our method, we conduct extensive experiments on two public domain adaptation datasets for point cloud classification and segmentation tasks. Empirical experiment results show that our CDND achieves state-of-the-art performance by a noticeable margin over existing approaches.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is **the challenges of Unsupervised Domain Adaptation (UDA) in point - cloud data**. Specifically, the author focuses on how to effectively bridge the gap between the source domain and the target domain (i.e., "domain gap") to reduce the need for a large amount of manually - labeled data during the training of deep networks. ### Main problems: 1. **Domain gap**: In practical applications, test data may come from a distribution different from that of the training data, which is called "domain gap". This gap causes a performance degradation when directly applying the model of the source domain to the target domain. 2. **Special characteristics of point - cloud data**: 3D point - cloud data is irregular, unstructured, and unordered. These characteristics exacerbate geometric changes, making existing 2D image UDA methods difficult to be directly applied to point - cloud data. ### Solutions: To solve the above problems, the author proposes the **Curvature Diversity - Driven Nuclear - Norm Wasserstein Domain Alignment (CDND)** method. This method mainly contains two innovative points: 1. **Curvature Diversity - driven Deformation Reconstruction (CurvRec)**: - By evaluating the curvature diversity of different regions of the point cloud, select regions with less information for deformation, thereby avoiding destroying semantically rich regions. - Use entropy to measure curvature diversity to ensure that the model can extract features from the most informative regions. 2. **Deformation - based Nuclear - norm Wasserstein Discrepancy (D - NWD)**: - When aligning the source domain and the target domain, not only consider the original samples but also the deformed samples, thereby creating a richer and more robust feature space. - Theoretical analysis shows that D - NWD can effectively reduce the distribution difference between the source domain and the target domain and is applicable to any deformation method. ### Experimental verification: To verify the effectiveness of CDND, the author conducted extensive experiments on two publicly available domain - adaptation datasets, namely the datasets for point - cloud classification and segmentation tasks. The experimental results show that CDND significantly outperforms existing methods on multiple metrics and achieves state - of - the - art performance. ### Formula summary: - **Curvature calculation formula**: \[ c=\frac{|\lambda_{\min}|}{\sum_{i = 1}^{K}|\lambda_i|} \] where \(\lambda_{\min}\) is the minimum eigenvalue and \(K\) is the number of eigenvalues. - **Curvature diversity (entropy) calculation formula**: \[ H(c_{\text{norm}})=-\sum_{i = 1}^{N_R}c_{i,\text{norm}}\cdot\log(c_{i,\text{norm}}+ 1\times10^{-10}) \] - **D - NWD objective function**: \[ W_N(\nu_s\cup\nu_{sd},\nu_t\cup\nu_{td})=\sup_{\|C\|_*\leq K}\mathbb{E}_{f_s\sim\nu_s\cup\nu_{sd}}[\|C(\hat{f}_s)\|_*]-\mathbb{E}_{f_t\sim\nu_t\cup\nu_{td}}[\|C(\hat{f}_t)\|_*] \] Through these methods, the author has successfully solved the key problems in point - cloud UDA and demonstrated the superior performance of his method in practical applications.

Curvature Diversity-Driven Deformation and Domain Alignment for Point Cloud

Domain Adaptation on Point Clouds Via Geometry-Aware Implicits

Self-Supervised Global-Local Structure Modeling for Point Cloud Domain Adaptation with Reliable Voted Pseudo Labels

Attention-based Cross-Layer Domain Alignment for Unsupervised Domain Adaptation

A New Bidirectional Unsupervised Domain Adaptation Segmentation Framework

Deformation depth decoupling network for point cloud domain adaptation

Point-Based Multilevel Domain Adaptation for Point Cloud Segmentation

Synergizing Contrastive Learning and Optimal Transport for 3D Point Cloud Domain Adaptation

Self-Supervised Boundary Point Prediction Task for Point Cloud Domain Adaptation

Bridging Domain Gap of Point Cloud Representations via Self-Supervised Geometric Augmentation

Cross-Dataset Point Cloud Recognition Using Deep-Shallow Domain Adaptation Network.

A Learnable Self-supervised Task for Unsupervised Domain Adaptation on Point Clouds

Progressive Classifier and Feature Extractor Adaptation for Unsupervised Domain Adaptation on Point Clouds

Masked Local-Global Representation Learning for 3D Point Cloud Domain Adaptation

Joint Distribution Adaptive-Alignment for Cross-Domain Segmentation of High-Resolution Remote Sensing Images

Deep Covariance Alignment for Domain Adaptive Remote Sensing Image Segmentation

Learning to Adapt SAM for Segmenting Cross-domain Point Clouds

AWCD: An Efficient Point Cloud Processing Approach via Wasserstein Curvature

Aligning Domain-Specific Distribution and Classifier for Cross-Domain Classification from Multiple Sources

PointDAN: A Multi-Scale 3D Domain Adaption Network for Point Cloud Representation

Domain Adaptive Sampling for Cross-Domain Point Cloud Recognition