3D Point Cloud Network Pruning: When Some Weights Do not Matter

TL;DR

The paper tackles the computational burden of 3D Point Cloud Neural Networks (PCNNs) by applying the Lottery Ticket Hypothesis to uncover highly sparse subnetworks that retain accuracy. Using magnitude-based pruning with iterative IMP and one-shot global pruning, it identifies winning tickets in PointNet, DGCNN, and PointCNN that achieve up to $99$\% sparsity while preserving performance on ModelNet40, ScanObjectNN, and ShapeNetCore; crucially, preserving the top $p$\% of weights suffices for accuracy, with base models attaining $87.5\%$ on ModelNet40 and sparse subnetworks reaching $86.8\%$ with just $1$\% of weights. The results show that most indispensable weights reside in convolutional layers, enabling substantial FC-layer pruning without harming accuracy and revealing that these winning tickets transfer across related 3D tasks. This work advances PCNN compression, offering pathways for efficient edge deployment and informing design choices that favor convolutional representations in 3D learning.

Abstract

A point cloud is a crucial geometric data structure utilized in numerous applications. The adoption of deep neural networks referred to as Point Cloud Neural Networks (PC- NNs), for processing 3D point clouds, has significantly advanced fields that rely on 3D geometric data to enhance the efficiency of tasks. Expanding the size of both neural network models and 3D point clouds introduces significant challenges in minimizing computational and memory requirements. This is essential for meeting the demanding requirements of real-world applications, which prioritize minimal energy consumption and low latency. Therefore, investigating redundancy in PCNNs is crucial yet challenging due to their sensitivity to parameters. Additionally, traditional pruning methods face difficulties as these networks rely heavily on weights and points. Nonetheless, our research reveals a promising phenomenon that could refine standard PCNN pruning techniques. Our findings suggest that preserving only the top p% of the highest magnitude weights is crucial for accuracy preservation. For example, pruning 99% of the weights from the PointNet model still results in accuracy close to the base level. Specifically, in the ModelNet40 dataset, where the base accuracy with the PointNet model was 87. 5%, preserving only 1% of the weights still achieves an accuracy of 86.8%. Codes are available in: https://github.com/apurba-nsu-rnd-lab/PCNN_Pruning

Figures (3)

• Figure 1: This paper deals with pruning 3D Point Cloud Neural Networks (PCNNs) for faster inference, especially on edge devices. (a) In 3D point cloud literature, traditionally, researchers carefully design less parametric deep models for this purpose (e.g., Spherical CNNs esteves2018learning, Dense Point liu2019densepoint, KCNet shen2018mining, Point PN zhang2023parameter). (b) In contrast, we iteratively prune popular PCNNs (PointCNN li2018pointcnn, DGCNN wang2019dynamic, PointConv wu2019pointconv) to produce task-specific subnetworks outperforming over-parameterized models. (c) Our results demonstrate the capability of the LTH to extract highly sparse subnetworks, termed "winning tickets," from an over-parameterized model. These winning tickets can achieve up to 99% sparsity, retaining only 1% of the original weights while still attaining accuracy levels comparable to the base model.
• Figure 2: The performance of lottery tickets discovered for 3D classification tasks across three model architectures: PointNet, DGCNN, and PointCNN, which handle point cloud data differently. Task-specific Iterative Magnitude Pruning (IMP) Global pruning outperforms the overparametrized baseline (0% pruned) and IMP Local pruning by a wide margin. For efficiency, one-shot global pruning at 99% sparsity exceeds all other methods evaluated. Each task's IMP method iteratively prunes 20% weights to identify the winning tickets or subnetworks, achieving competitive or superior performance to the over parametrized models with reduced computation and parameters.
• Figure 3: The first two figures from the left illustrate the impact of global pruning on models such as PointNet and DGCNN. Based on the weight magnitude (importance), LTH prunes some Conv. layers with a significantly lower proportion of weights than FC layers. It tells LTH to automatically identify essential weights related to geometric transformations and other structural constraints of 3D data. The table on the right shows an ablation study for pruning the Conv. and FC layers separately.