3D Point Cloud Densification for LiDAR Domain Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing 3D object detection networks trained on unpaired datasets from different LiDAR sensors with varying resolutions struggle to generalize due to domain shift, leading to inefficiencies in training and performance.
Innovation Solution
A method for domain adaptation through data densification, upsampling sparse labeled source-domain datasets to match the resolution of target-domain datasets, using a trained domain adaptation network to generate encoded and reconstructed 3D point clouds, and interpolating data points to create a densified 3D point cloud.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If a neural network is trained on source-domain data from a LiDAR sensor with different resolution, then training data availability is improved, but generalization performance deteriorates due to domain shift
Solution Approach 1:
A domain adaptation network is introduced as an intermediary component between the source-domain LiDAR data and the target-domain processing pipeline. This network transforms the source-domain point cloud data into a representation that matches the target-domain resolution characteristics, enabling effective training while maintaining generalization performance across different sensor domains
Solution Approach 2:
The patent applies parameter changes by modifying the resolution characteristics of the source-domain point cloud data through the domain adaptation network. The network learns to transform spatial parameters and density parameters of the point cloud to match the target domain's resolution profile, thereby resolving the domain shift issue while utilizing abundant source-domain training data
2Adaptability or versatility
If source-domain point cloud data is used for training, then data diversity is improved, but resolution mismatch with target-domain data worsens
Solution Approach 1:
The domain adaptation network serves as a mediator that receives diverse source-domain point cloud data and transforms it into a format with resolution characteristics matching the target domain. This intermediary transformation preserves the diversity benefits while achieving precise resolution alignment through learned spatial transformations
Solution Approach 2:
The patent addresses resolution mismatch by operating in the feature space dimension rather than directly in physical space. The domain adaptation network learns transformations in this abstract dimension that effectively align the resolution characteristics of source and target domains, enabling diverse data utilization with precise matching
Data Source
AI summary
Devices, systems, methods, and media are disclosed for domain adaptation using data densification. Example embodiments described herein receives LiDAR 3D point clouds from a source-domain and introduces interpolated 3D points inferred by a trained deep learning neural network to output a denser version of the input 3D point cloud with increased resolution. The trained domain adaptation network reconstructs the source-domain 3D point cloud data, generates translation vectors to compute interpolated 3D point cloud data and merges the reconstructed 3D point cloud data and the interpolated 3D point cloud data to output a densified 3D point cloud resembling data 3D point clouds captured generated by the target LiDAR sensor from the source-domain.