RW-Net: Enhancing Few-Shot Point Cloud Classification with a Wavelet Transform Projection-based Network

TL;DR

RW-Net addresses the scarcity of labeled data in 3D few-shot classification by fusing Rate-Distortion Explanation (RDE) with wavelet transforms inside a projection-based backbone. It extends CartoonX to the wavelet domain and integrates Wavelet Attention Blocks within a ViewNet-like pipeline, using a masked, multi-view projection strategy to focus on salient, low-frequency geometric information. The approach yields consistent state-of-the-art performance on ModelNet40, ModelNet40-C, and ScanObjectNN, while demonstrating robustness to noise and occlusions and providing insights into component contributions via comprehensive ablations. This work indicates that explainability-driven feature distillation and frequency-domain filtering can significantly improve learning efficiency and generalization in low-data regimes for 3D vision tasks.

Abstract

In the domain of 3D object classification, a fundamental challenge lies in addressing the scarcity of labeled data, which limits the applicability of traditional data-intensive learning paradigms. This challenge is particularly pronounced in few-shot learning scenarios, where the objective is to achieve robust generalization from minimal annotated samples. To overcome these limitations, it is crucial to identify and leverage the most salient and discriminative features of 3D objects, thereby enhancing learning efficiency and reducing dependency on large-scale labeled datasets. This work introduces RW-Net, a novel framework designed to address the challenges above by integrating Rate-Distortion Explanation (RDE) and wavelet transform into a state-of-the-art projection-based 3D object classification architecture. The proposed method capitalizes on RDE to extract critical features by identifying and preserving the most informative data components while reducing redundancy. This process ensures the retention of essential information for effective decision-making, optimizing the model's ability to learn from limited data. Complementing RDE, incorporating the wavelet transform further enhances the framework's capability to generalize in low-data regimes. By emphasizing low-frequency components of the input data, the wavelet transform captures fundamental geometric and structural attributes of 3D objects. These attributes are instrumental in mitigating overfitting and improving the robustness of the learned representations across diverse tasks and domains. To validate the effectiveness of our RW-Net, we conduct extensive experiments on three datasets: ModelNet40, ModelNet40-C, and ScanObjectNN for few-shot 3D object classification. The results demonstrate that our approach achieves state-of-the-art performance and exhibits superior generalization and robustness in few-shot learning scenarios.

Figures (5)

• Figure 1: Left: The workflow of CartoonX kolek2022cartoon where $x$ is an input image and $s$ is a continuous mask initialized to be all ones. Right: Our method jointly optimizes image mask and classification model parameters. The symbol $c$ represents a given point cloud, $x$ denotes (one of) the projected images from $c$, and $\odot$ denotes Hadamard product.
• Figure 2: The structure of our RW-Net consisting of model $Q$ (and an identical model $\hat{Q}$) which is built on ViewNet chen2023viewnet. Note that we replace the conventional convolutional block with a wavelet attention block to efficiently capture the semantic contents of the input features while suppressing noise. The symbol $\oplus$ denotes element-wise addition.
• Figure 3: Structure of wavelet attention block. $\mathbf{I}_{ll}$, $\mathbf{I}_{lh}$, and $\mathbf{I}_{hl}$ adapted from zhao2022wavelet. The symbol $\oplus$ denotes element-wise addition, and $\odot$ denotes element-wise multiplication.
• Figure 4: Comparison between unprocessed projected images (top) and cartoonX processed images (bottom).
• Figure 5: Visualization of wavelet transform extracting different frequency components of the original images.