3D Reconstruction Using Normalized Object Location Fields
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing three-dimensional modeling technologies face challenges in efficiently and accurately obtaining real and available three-dimensional digital data at low costs, due to the high cost of high-precision laser scanners and the low accuracy of conventional image technology in model retrieval.
Innovation Solution
A three-dimensional reconstruction method using a two-dimensional image, which involves obtaining an image of an object, determining a normalized object location field (NOLF) image using a deep learning network, retrieving a corresponding three-dimensional model from a database, and performing three-dimensional reconstruction based on the model and camera pose.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a high-precision laser scanner is used to obtain three-dimensional point cloud information, then measurement precision is improved, but device cost increases
Solution Approach 1:
The patent uses a two-dimensional image as a copy or representation of the three-dimensional object, combined with a pre-stored three-dimensional model, to reconstruct the object's three-dimensional information without requiring expensive laser scanning equipment. The image serves as a proxy that, when processed with the model, yields accurate three-dimensional data at low cost.
Solution Approach 2:
The patent introduces a three-dimensional model as an intermediary between the two-dimensional image and the desired three-dimensional reconstruction. The model database stores pre-computed three-dimensional models that serve as mediators to bridge the gap between simple image input and accurate three-dimensional output, enabling high-precision reconstruction without direct three-dimensional sensing.
2Device complexity
If pre-constructed model database with projected images is used, then device complexity is reduced, but measurement precision deteriorates
Solution Approach 1:
The patent transforms the retrieval problem from two-dimensional image matching to three-dimensional spatial relationship matching. By storing three-dimensional models with their spatial relationships and using depth information from the image, the system performs matching in an enhanced three-dimensional feature space, improving accuracy while maintaining computational efficiency.
Solution Approach 2:
The patent changes the matching parameters from traditional two-dimensional image similarity metrics to three-dimensional spatial parameters including depth, position, and orientation. This parameter transformation enables the system to accurately match objects based on their true three-dimensional characteristics rather than just their two-dimensional projections, significantly improving retrieval accuracy.
3Ease of manufacture
If preset three-dimensional model is projected at preset location and angle, then ease of manufacture is improved, but adaptability deteriorates
Solution Approach 1:
The patent makes the model retrieval and matching process dynamic by adjusting the projection angle and position parameters based on the actual image content and camera pose. Instead of using fixed preset projections, the system dynamically computes the optimal projection parameters that match the captured image, enabling adaptation to various angles and positions while maintaining ease of model preparation.
Data Source
AI summary
An embodiment of this application discloses a three-dimensional reconstruction method. The method in this embodiment of this application includes: obtaining an image of a first object and a camera pose of the image; determining a first normalized object location field NOLF image of the first object in the image by using a first deep learning network, where the first NOLF image indicates a normalized three-dimensional point cloud of the first object at a photographing angle of view of the image; determining, from a plurality of three-dimensional models in a model database based on the first NOLF image, a first model corresponding to the first object; determining a pose of the first object based on the first model and the camera pose of the image; and performing three-dimensional reconstruction on the first object in the image based on the first model and the pose of the first object.