Point Cloud Upsampling via Occupancy Map Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current point cloud compression techniques suffer from quantization loss, resulting in a coarser sub-sampled representation of point clouds, which leads to a lower Level-of-Detail (LoD) after data transmission or storage, and there is a need to upsample point cloud data to minimize this loss and achieve a higher LoD, especially in applications like augmented reality (AR) and telepresence where detailed representation is required.
Innovation Solution
A method and system for point cloud processing that involves receiving quantized point cloud data, generating an occupancy map for lost voxels during quantization, and reconstructing the point cloud by populating these voxels, using a U-net architecture to learn the underlying 3D structure and generate a denser point cloud with improved detail, enabling display adaptation by upsampling the point cloud to a higher LoD.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If point cloud compression is applied to reduce data size, then data transmission and storage efficiency is improved, but quantization loss increases resulting in coarser representation
Solution Approach 1:
The system performs preliminary upsampling of the quantized point cloud before the actual processing or display. By anticipating the need for higher detail and pre-upsampling the compressed data, the system prepares a higher-resolution representation without requiring the entire original high-resolution data to be stored or transmitted, thus resolving the contradiction between data size and representation detail.
Solution Approach 2:
The system dynamically adjusts the point cloud resolution based on the application requirements. Instead of using a fixed resolution, the system can switch between compressed low-resolution and upsampled high-resolution representations depending on the task needs, allowing flexible adaptation between data efficiency and detail requirements.
2Productivity
If point cloud data is downsampled to reduce computational complexity, then processing speed is improved, but Level-of-Detail decreases
Solution Approach 1:
The system pre-upsamples the downsampled point cloud data before processing or display operations. By performing the upsampling action in advance on the compressed data, the system prepares high-detail representations without the need to process the entire original high-resolution data set, thus maintaining both processing speed and detail level.
Solution Approach 2:
The system creates a copy of the downsampled point cloud and applies upsampling algorithms to generate a higher-resolution version. This copying approach allows the system to work with the original low-resolution data for efficiency while generating detailed representations only when needed, resolving the contradiction between processing speed and detail.
3Quantity of substance
If quantization step size is increased to reduce data precision requirements, then compression ratio is improved, but quantization loss increases
Solution Approach 1:
The system accepts the harmful effect of quantization loss as a given and uses it to drive the upsampling process. By recognizing that quantization inevitably loses detail, the system deliberately designs the upsampling algorithm to compensate for this loss by learning the underlying 3D structure from the quantized data and generating corrected high-resolution representations, thus converting the harmful quantization effect into a catalyst for improving representation quality.
Solution Approach 2:
The upsampling process incorporates feedback mechanisms where the system continuously refines the reconstructed point cloud by comparing it with the original quantized data and adjusting the upsampled representation accordingly. This feedback loop allows the system to minimize quantization loss by iteratively improving the reconstruction accuracy, thus resolving the contradiction between compression ratio and information loss.
Data Source
AI summary
A method, computer program, and computer system is provided for processing point cloud data. Quantized point cloud data including a plurality of voxels is received. An occupancy map is generated for the quantized point cloud corresponding to lost voxels during quantization from among the plurality of voxels. A point cloud is reconstructed from the quantized point cloud data based on populating the lost voxels.