A Semantic Uncertainty Aware Method for Point Clouds Based on Neighborhood Aggregation Monte Carlo Deactivation

Abstract

The present invention proposes a point cloud semantic uncertainty perception method based on neighborhood aggregation Monte Carlo deactivation, which includes obtaining the original point cloud of the scene to be processed; using the original point cloud as input, and using PointNet(++) as The basic model constructs the NSA-MC-dropout framework, in which, in the encoding stage, the NSA-MC-dropout framework generates feature vectors of different granularities. After the perceptron performs inference, it generates random inference results for each unordered point; realizes a single random inference by fusing the random inference results to establish the prediction distribution of each unordered point in the original point cloud; quantifies by capturing the amount of information contained in the prediction distribution Point Cloud Semantic Uncertainty or Quantify point cloud semantic uncertainty by modeling the degree of dispersion of the predicted distribution. A framework for uncertainty-aware point cloud semantic segmentation is implemented without increasing model parameters and inference times.

Description

Point Cloud Semantic Uncertainty Awareness Based on Neighborhood Aggregation Monte Carlo Deactivation technical field [0001] The present invention belongs to the field of three-dimensional visual pattern recognition. Background technique [0002] Scenarios such as robot grasping, path planning and automatic driving require the use of 3D laser scanning equipment to detect surrounding scenes. Perform 3D modeling to form point cloud data (environmental structure topology data composed of massive discrete points), point cloud semantic segmentation (recognition Identifying the semantic label of each point in the point cloud) is the basis for the machine to make further decisions. However, the current semantic segmentation technology It is difficult to accurately identify the label of each point in the point cloud. Misidentification creates risk for machine decision-making. therefore, It is necessary to use the point cloud semantic segmentation technology with uncertainty...

AI Technical Summary

Problems solved by technology

However, whatever uncertainty component is used for the optimization of the model, it must be established from the predictive distribution

However, the traditional method of establishing the forecast distribution relies on the introduction of MC-dropout or additional variance parameters, and this inefficient method is difficult to be promoted

[0006] To sum up, the existing uncertainty assessment and uncertainty-guided learning methods all have the problem of time-consuming establishment of the target prediction distribution.

For point clouds, which contain massive data containing millions or even tens of millions of points, traditional predictive distribution establishment methods are difficult to apply

Images