Test-Time Adaptation in Point Clouds: Leveraging Sampling Variation with Weight Averaging

TL;DR

The paper tackles distribution shifts in 3D point cloud classification by proposing a fully source-free test-time adaptation method. It introduces sampling variation through FPS and KNN to create multiple input representations, adapts normalization layers per representation via entropy-minimizing TENT, and then aggregates the adapted weights by averaging to yield a robust final model. The approach aligns with flat-minima ideas and robust risk minimization, offering parallelizable adaptations and maintaining negligible memory overhead. Empirically, SVWA TTA delivers state-of-the-art performance across ModelNet-40C, ShapeNet-C, and ScanObjectNN-C with backbones Point-MAE, PointNet, and DGCNN, demonstrating strong generalization under real-world corruptions while remaining efficient.

Abstract

Test-Time Adaptation (TTA) addresses distribution shifts during testing by adapting a pretrained model without access to source data. In this work, we propose a novel TTA approach for 3D point cloud classification, combining sampling variation with weight averaging. Our method leverages Farthest Point Sampling (FPS) and K-Nearest Neighbors (KNN) to create multiple point cloud representations, adapting the model for each variation using the TENT algorithm. The final model parameters are obtained by averaging the adapted weights, leading to improved robustness against distribution shifts. Extensive experiments on ModelNet40-C, ShapeNet-C, and ScanObjectNN-C datasets, with different backbones (Point-MAE, PointNet, DGCNN), demonstrate that our approach consistently outperforms existing methods while maintaining minimal resource overhead. The proposed method effectively enhances model generalization and stability in challenging real-world conditions.

Figures (8)

• Figure 1: Overview of our 3D methodology. First, is applied to generate different samplings from the input point cloud. Patchification is then performed using for patch centers and to form patches (a and b). The Normalization Layer (NL) weights are adapted using the TENT algorithm for each sampling. Finally, weight averaging is applied across all adapted weights to enhance robustness and generalization.
• Figure 2: Impact of Different $N^V$ on Model Accuracy
• Figure 3: Comparison between Sampling Variation and Different Augmentations
• Figure 4: Impact of Batch Size on Accuracy for Two Methods
• Figure 5: Impact of Iteration on Accuracy for Two Methods