Hybrid Method for Point Cloud Classification

doi:10.1109/ACCESS.2025.3525739

2025-01-01

Abdurrahman Hazer *

Remzi Yıldırım

DOI：10.1109/ACCESS.2025.3525739

原文链接

原文求助

摘要

En 中文

In this study, we introduce a novel hybrid ResPOL method that integrates a bidirectional architecture with a newly developed Patch-Offset Lambda (POL) mechanism for feature extraction from point cloud data. This hybrid approach effectively combines the Residual Machine Learning Perceptron (ResMLP) with the POL mechanism to capture both local and global features of 3D point clouds. The ResMLP component hierarchically extracts high-frequency local features from point cloud patches, while the POL mechanism is specifically designed to capture low-frequency global features. This dual extraction process ensures minimal loss of both high and low-frequency features. The POL mechanism employs Lambda layers within a linear framework, significantly enhancing the classification speed and accuracy compared to traditional attention mechanisms that suffer from quadratic complexity and non-linear structures. By processing local and global features in parallel, the Hybrid ResPOL method combines these features and feeds them into the classification head, optimizing performance. Experimental results indicate that the Hybrid ResPOL method achieves an overall accuracy (OA) of 94.3% and a mean accuracy (mAcc) of 91.3% on the ModelNet40 dataset. Additionally, it demonstrates robust performance on the challenging ScanObjectNN dataset, with accuracies of 92% OA and 91.2% mAcc. The method processes data at rates of 205.1 samples per second during training and 493.6 samples per second during testing, outperforming the PointMLP method by a factor of 4.3. The superior performance of Hybrid ResPOL in both accuracy and speed highlights its effectiveness over existing attention-based methods.

Keyword:

Point cloud classification

patch offset lambda

linear attention

deep learning

3D computer vision

Point cloud classification

patch offset lambda

linear attention

deep learning

3D computer vision