A point cloud self-learning network based on contrastive learning for classification and segmentation

Haoran Zhou

doi:10.1007/s00371-023-03248-4

A point cloud self-learning network based on contrastive learning for classification and segmentation

2024-01-23

0

PRE

AI

DOI：10.1007/s00371-023-03248-4

原文链接

原文求助

摘要

En 中文

In the field of point cloud representation learning, many self-supervised learning methods aim to address the issue of conventional supervised learning methods relying heavily on labeled data. Particularly in recent years, contrastive learning-based methods have gained an increasing popularity. However, most of the current contrastive learning methods solely rely on conventional random augmentation, limiting the effectiveness of representation learning. Moreover, to prevent model collapse, they construct positive and negative sample pairs or explicit clustering centers, which adds complexity to data preprocessing operations. To address these challenges effectively and achieve accurate point cloud classification and segmentation, we propose PointSL, a self-learning network for point clouds based on contrastive learning. PointSL incorporates a learnable point cloud augmentation (LPA) module, which transforms samples with high precision, significantly improving the augmentation effect. To further enhance feature discrimination, PointSL introduces a self-learning process along a refined feature predictor (FFP). This innovative approach leverages the attention mechanism to facilitate mutual feature prediction between pairs of point clouds, thereby continuously improving discriminant performance. Additionally, the network constructed a simple yet effective self-adaptive loss function that optimizes the entire network through gradient feedback. For pretraining, it is beneficial to obtain encoders with a better generalization and a higher accuracy. We evaluate PointSL on benchmark datasets such as ModelNet40, Sydney Urban Objects and ShapeNetPart. Experimental results demonstrate that PointSL outperforms state-of-the-art self-supervised methods and supervised counterparts, achieving exceptional performance in classification and segmentation tasks. Notably, on the Sydney Urban Objects and ModelNet40 datasets, PointSL achieves OA and AA metrics of 80.6%, 69.9%, 94.2% and 91.4%, respectively. On the ShapeNetPart dataset, PointSL achieves Inst.mIoU and Cls.mIoU metrics of 86.3% and 85.1%, respectively.

Keyword:

Contrastive learning

Point cloud representation learning

Data augmentation

A point cloud self-learning network based on contrastive learning for classification and segmentation

摘要

期刊

机构

引用论文