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

TL;DR

This work introduces Progressive Classifier and Feature Extractor Adaptation (PCFEA) for unsupervised domain adaptation on point clouds by tightly coupling classifier and feature extractor updates. It combines a macro-level Progressive Target-Styled Feature Augmentation (PTFA) with a micro-level Intermediate Domain Feature Extractor Adaptation (IDFA), enabling gradual, mutually beneficial adaptation toward the target domain. The method constructs a sequence of intermediate domains and optimizes both classifier and feature extractor through stage-wise augmentation and alignment losses, achieving state-of-the-art results on PointDA-10 and GraspNetPC-10. The approach addresses target-distribution estimation challenges and target-sample variability in 3D UDA, offering a practical framework with strong empirical gains and accessible code.

Abstract

Unsupervised domain adaptation (UDA) is a critical challenge in the field of point cloud analysis. Previous works tackle the problem either by feature extractor adaptation to enable a shared classifier to distinguish domain-invariant features, or by classifier adaptation to evolve the classifier to recognize target-styled source features to increase its adaptation ability. However, by learning domain-invariant features, feature extractor adaptation methods fail to encode semantically meaningful target-specific information, while classifier adaptation methods rely heavily on the accurate estimation of the target distribution. In this work, we propose a novel framework that deeply couples the classifier and feature extractor adaption for 3D UDA, dubbed Progressive Classifier and Feature Extractor Adaptation (PCFEA). Our PCFEA conducts 3D UDA from two distinct perspectives: macro and micro levels. On the macro level, we propose a progressive target-styled feature augmentation (PTFA) that establishes a series of intermediate domains to enable the model to progressively adapt to the target domain. Throughout this process, the source classifier is evolved to recognize target-styled source features (\ie, classifier adaptation). On the micro level, we develop an intermediate domain feature extractor adaptation (IDFA) that performs a compact feature alignment to encourage the target-styled feature extraction gradually. In this way, PTFA and IDFA can mutually benefit each other: IDFA contributes to the distribution estimation of PTFA while PTFA constructs smoother intermediate domains to encourage an accurate feature alignment of IDFA. We validate our method on popular benchmark datasets, where our method achieves new state-of-the-art performance. Our code is available at https://github.com/xiaoyao3302/PCFEA.

Figures (4)

• Figure 1: Illustration of our proposed progressive classifier and feature extractor adaptation (PCFEA) approach using samples from the same category in different domains. (a) The gaps between different target samples to the source domain are different, while a vanilla model trained solely with source data cannot recognize certain target samples with a large gap to the source domain. (b) A model trained with augmented source samples will improve the generalization of the model. (c) Considering the uniqueness of each sample, we select part of the source and target samples, i.e., the dark-colored solid ones, to progressively generate new intermediate domains toward the target domain. We construct feature augmentations and perform feature alignment towards the new intermediate domain and the model trained will gradually approach the target domain.
• Figure 2: Visualization of the changing process of the distributions of the source, the target, and different intermediate domains. We use samples from the same category under the adaptation scenario ModelNet-10 $\rightarrow$ ScanNet-10 as an example. $D_k$ indicates the constructed intermediate domain during the $k$-th stage.
• Figure 3: An overview of our proposed method for unsupervised domain adaptation on point clouds. We progressively generate intermediate domains that gradually resemble the target domain. Taking one intermediate domain as an example, we first propose a progressive target-styled feature augmentation (PTFA) approach to construct semantically meaningful feature augmentations on the source features extracted by $\varphi_f$ towards the intermediate domain, which is then fed into the classifier $\varphi_{cls}$ to generate source predictions. The source predictions are supervised by the ground-truth labels $y_s$. Directly optimizing the corresponding $\mathcal{L}_{PTFA}$ enables the classifier to be adapted toward the intermediate domain. We further propose an intermediate domain feature extractor adaptation (IDFA) strategy to encourage the features extracted from the source samples and the target samples to approach the estimated mean of the intermediate domain using $\mathcal{L}_{IDFA}$.
• Figure 4: Ablation study on the sensitivity of our method to the training epoch at each intermediate domain construction stage, i.e., $\tau$. Experiments are conducted on the PointDA-10 dataset. We report the average recognition accuracy on the target domain and the average training duration.