A semantic segmentation method and system for a point cloud encoder

By constructing a dual-branch point cloud knowledge distillation network framework, prior information is generated and student models are trained, solving the problem of low accuracy of point cloud semantic segmentation models on resource-constrained devices and achieving efficient semantic segmentation results.

CN122223341APending Publication Date: 2026-06-16TIANFU JIANGXI LAB
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
TIANFU JIANGXI LAB
Filing Date
2026-05-18
Publication Date
2026-06-16

AI Technical Summary

Technical Problem

In existing technologies, point cloud semantic segmentation models struggle to maintain high accuracy while reducing inference costs, especially on resource-constrained edge devices where semantic segmentation accuracy remains low.

Method used

A bi-branch point cloud knowledge distillation network framework comprising a teacher model and a student model is constructed to generate local geometric inductive bias, multi-scale geometric topological constraints, gradient saliency distillation information, and knowledge cache reuse information. The student model is trained by jointly optimizing the objective, thereby reducing computational complexity and improving semantic segmentation accuracy.

Benefits of technology

While maintaining end-to-end direct reasoning capabilities, it effectively reduces the computational complexity and memory overhead of the student model, improves semantic segmentation accuracy, and adapts to resource-constrained edge subject perception application scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122223341A_ABST
    Figure CN122223341A_ABST
Patent Text Reader

Abstract

The embodiment of the application provides a kind of semantic segmentation method and system for point cloud encoder, it is related to three-dimensional point cloud processing and computer vision technical field.The method is after obtaining point cloud data, by constructing the double-branch point cloud knowledge distillation network framework including teacher model and student model, generate prior information.The prior information includes local geometric inductive bias, multiscale geometric topological constraint, gradient saliency distillation information and knowledge cache reuse information.Again based on prior information, construct joint optimization target, and according to joint optimization target, student model is trained using point cloud data.The method can effectively reduce the computational complexity and memory overhead of student model while maintaining the end-to-end direct inference capability of the original point cloud, by jointly pruning the network depth, channel width and attention head number, improve the semantic segmentation accuracy under limited resources.
Need to check novelty before this filing date? Find Prior Art