Point Cloud Upsampling Through Dual Back-Projection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Obtaining high-resolution point clouds is expensive and burdensome due to the cost and storage requirements of dense point clouds, while sparse point clouds result in rough images that fail to capture object details, leading to poor visual experience.
Innovation Solution
A method involving a dual back-projection process for iteratively upsampling and downsampling point clouds using feature and space domains to increase density, employing a trained network model and noise addition for robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If dense point clouds are obtained using measurement instruments, then the density and detail of the point cloud increases, but the acquisition cost and storage requirements increase significantly
Solution Approach 1:
The patent uses a learned mapping model to generate a dense point cloud as a copy from a sparse point cloud, avoiding the need to physically acquire the dense point cloud through expensive measurement instruments. The model learns the transformation from sparse to dense representation and applies it to generate realistic dense point clouds synthetically
Solution Approach 2:
The patent replaces the mechanical measurement system (laser scanners, coordinate measuring machines) with a computational learning system. Instead of physically scanning objects to obtain dense point clouds, the system uses trained neural networks to synthesize dense point clouds from sparse inputs, substituting physical measurement with computational generation
2Quantity of substance
If dense point clouds are obtained using measurement instruments, then the density and detail of the point cloud increases, but the storage and calculation requirements increase
Solution Approach 1:
The patent performs preliminary learning during a training phase where the mapping model is trained on pairs of sparse and dense point clouds. This preliminary action stores the learned transformation patterns in the model parameters, so that during actual use, only the lightweight inference process is needed, avoiding the need to store multiple dense point cloud datasets
Solution Approach 2:
The model creates a computational copy of the dense point cloud structure through learned transformation rather than storing actual dense point cloud data. This allows the system to work with dense representations during inference while maintaining efficient storage requirements, as only the model parameters need to be stored
3Ease of manufacture
If sparse point clouds are used, then the acquisition cost and storage requirements are reduced, but the visual quality and detail representation deteriorate
Solution Approach 1:
The patent generates a high-quality copy of the dense point cloud from the sparse input through the learned mapping model. This copied dense representation restores the visual quality and detail that would otherwise be lost in sparse point clouds, while maintaining the low acquisition cost of using sparse inputs
Solution Approach 2:
The patent transforms the point cloud representation parameters through the learned mapping model, changing from sparse coordinates to dense coordinates. This parameter transformation enhances the visual quality and detail representation while maintaining the efficiency benefits of starting with sparse data
Data Source
AI summary
A method in an illustrative embodiment includes: obtaining a first point cloud based on an input point cloud, a point number of the first point cloud being greater than a point number of the input point cloud; obtaining a first group of point clouds based on the first point cloud, the first group of point clouds including a plurality of point clouds; obtaining a second group of point clouds based on the input point cloud and the first group of point clouds, the second group of point clouds including a plurality of point clouds; and obtaining a target point cloud based on the first point cloud and the second group of point clouds, a point number of the target point cloud being greater than the point number of the input point cloud.