Research On Feature Extraction of Point Cloud Data Based on Contrastive Learning

Abstract

Due to the development of laser radar, depth camera and other technologies, point cloud data is used in more and more fields. However, compared with two-dimensional image data, the cost of manually labeling point cloud data is higher. This paper present a simple contrastive process to obtain the feature extraction encoder of point cloud data through self-supervised learning, which can provide better support for tasks such as classification and segmentation. We translate a mini batch of date into two crops, the corresponding point clouds data are treated as positive example, and the not corresponding data are treated as negative example. Using InfoNCE as target function to get the unique feature of each data. Comparing to other existing contrastive structure, it performs a higher accuracy in classification task based on ModelNet40. At the same time, we used rotation, randomly cutting and randomly dropout point to realize data augmentation based on ModelNet40 for improving the performance of feature extraction.