PointADAM: Unsupervised Adversarial Domain Adaptation on Point Clouds With Metric Learning via Compact Feature Representation

Abstract

Domain adaptation can mitigate the problem of limited labels in deep learning training. Nevertheless, extending the 2D domain adaptation method directly to 3D encounters challenges unique to point clouds, frequently resulting in inadequate feature alignment and a lack of discriminative features for decision boundaries. In light of this, we propose an unsupervised adversarial domain adaptation with metric learning (PointADAM) via compact feature representation. PointADAM is a two-stage architecture. In the first stage, it extracts high-level semantic and low-level local features from source and target domains based on the geometric relations of point cloud neighbors. Subsequently, it aligns the features from the two domains in an unsupervised adversarial manner, thereby obtaining a shared common feature space. In the second stage, a self-training method is employed to label the target domain with pseudo-labels. It is followed by selecting target domain samples with high confidence. To enhance discriminative feature learning, joint supervision is employed through both cross-entropy loss and center loss. Cross-entropy is utilized to calculate the probability of each category for classification and center loss is utilized to compact intra-class distribution and disperse inter-class distance, thereby improving classification accuracy by optimizing the decision boundary. Experimental results on the PointDA-10 dataset illustrate that our method performs relatively better than the state-of-the-art methods for domain adaptation on the point cloud by 3.1% in accuracy.