.png){kind=icon}
.png){kind=icon}
Return
PointCore: An efficient framework for unsupervised point cloud anomaly detection using joint local-global features
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neunet.2025.108446.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• Propose a novel architecture named PointCore, which uses a single memory bank to store local (coordinate) and global (PointMAE) features, reducing computational cost and feature mismatch interference. • Introduce a ranking-based normalization method to eliminate the distribution differences of anomaly scores, and adopt the point-to-plane Iterative Closest Points algorithm to optimize the point cloud registration results, enhancing the robustness of decision-making. • Experiments on the Real3D-AD dataset show that compared with methods like Reg3D-AD, PointCore achieves competitive inference time and the best performance in both detection and localization. • Combine global-local registration and memory bank construction, comprehensively utilize point cloud information, improve the recall rate of anomaly detection and reduce the false positive rate, suitable for scenarios that require precise detection of defective samples.
Journal
IF:
6.3
Papers:
7.7K
Citations:
3.0W
Organization
No organization information available