Three-dimensional scene segmentation method and device under weak supervision

Abstract

The invention provides a three-dimensional scene segmentation method and device under weak supervision, and relates to the technical field of computer vision and machine learning, and the method comprises the steps of marking a point on each instance to give an instance position and a semantic category, and extending the marked point to an over-segmentation block, namely a block-level label, through an over-segmentation algorithm; building an over-segmentation block graph by taking each over-segmentation block as a node, and performing multi-level combination processing on the nodes in the over-segmentation block graph, so that labeling information is expanded to all over-segmentation blocks from labeled over-segmentation blocks, and pseudo labels are generated for all unlabeled points in a point cloud scene. Therefore, each instance in the three-dimensional scene is marked with a point to represent the position of the instance, relatively important information can be effectively marked with relatively low cost, the pseudo-label is generated for the unmarked point in the point cloud scene, and the generated pseudo-label can be directly used as supervision information to train an existing strong supervision point cloud scene segmentation network.

Description

Technical field [0001] This application relates to the field of computer vision and machine learning, in particular, involving a three-dimensional scenario segmentation method and apparatus under weak supervision. Background technique [0002] In recent years, significant progress has been made in analyzing different three-dimensional geometric data structures, including point clouds, mesh, voxels, multi-view and implicit functions. Due to the popularity of various scanning equipment, three-dimensional dot cloud data has become very easy to obtain, therefore caught more and more attention. Recently, many depth learning methods that have been running directly on the cloud have been proposed, and the encouraging performance is achieved. [0003] Point cloud instance segmentation and semantic segmentation is two basic but challenging tasks in the 3D understanding. Given a point cloud scene, instance split is intended to find all existing objects and use different instance tags and c...

AI Technical Summary

Problems solved by technology

[0005] For this reason, the first purpose of this application is to propose a 3D scene segmentation method under weak supervision to solve the problem of huge labeling cost of strong supervision, by marking a point on each instance to give the instance location and semantic category , use the over-segmentation algorithm to extend the labeling points to the over-segmentati...

Images