A point cloud affordance region segmentation method based on self-correcting cross attention learning

By employing a self-calibrating cross-attention learning method, the problem of feature alignment in point cloud availability region segmentation is solved, achieving higher segmentation accuracy and acquisition of contextual interaction information, thus improving the segmentation effect of point cloud availability regions.

CN118072021BActive Publication Date: 2026-07-03NANJING UNIV OF POSTS & TELECOMM
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
NANJING UNIV OF POSTS & TELECOMM
Filing Date
2024-03-14
Publication Date
2026-07-03

AI Technical Summary

Technical Problem

Existing technologies for point cloud availability region segmentation suffer from issues such as feature-background mismatch and feature-foreground-background entanglement during feature alignment, resulting in insufficient segmentation accuracy.

Method used

A self-correcting cross-attention learning method is adopted to align and fuse image features and point cloud features through local feature extraction, self-attention mechanism and cross-attention calculation. Combined with loss function to optimize the training process, the segmentation accuracy is improved.

Benefits of technology

It improves the segmentation accuracy of point cloud availability regions, solves the matching and entanglement problems in feature alignment, enhances the acquisition of contextual interaction information, and improves the segmentation effect.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118072021B_ABST
    Figure CN118072021B_ABST
Patent Text Reader

Abstract

The application discloses a point cloud availability region segmentation method based on self-correcting cross attention learning, and belongs to the technical field of point cloud classification and segmentation. The method is as follows: local feature extraction is performed on the obtained data; similarity calculation is performed on image features and point cloud features to obtain a weight matrix, and an enhanced feature of the image features and the point cloud features is obtained through a self-attention mechanism; self-correcting attention learning is used to align the image enhanced feature and the point cloud enhanced feature, and a joint feature of a high-structure-similarity region is obtained; related factors between an object and a background and between the object and a subject are jointly modeled to obtain an enhanced feature containing context information; and a loss function is established to optimize a training process. The application realizes the alignment of the high-structure-similarity region of the point cloud and the image by using interpolation and pooling to segment the image enhanced feature and the point cloud enhanced feature; the self-correcting attention learning method is used to solve the problems of feature background mismatching and feature foreground and background entanglement in feature alignment, and the accuracy of point cloud availability region segmentation is improved.
Need to check novelty before this filing date? Find Prior Art