Satellite 3D Reconstruction with ML Depth Maps and Camera Parameters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for 3D reconstruction from satellite imagery are inefficient and lack accuracy due to lower resolution and time constraints, especially when capturing large areas.
Innovation Solution
A method using deep learning with a trained Machine Learning Network (MLN) that predicts depth maps from overlapping satellite images, utilizing higher-resolution ground truth data from aircraft imagery for training, and incorporates imaging device parameters to enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If satellite images are used for 3D reconstruction, then large areas can be covered efficiently, but the resolution is lower compared to aerial images
Solution Approach 1:
The patent creates synthetic training data by rendering satellite images and depth maps from 3D surface models. This copying approach allows the MLN to be trained on high-quality synthesized data that mimics real satellite imagery, enabling the network to learn accurate depth estimation from lower-resolution satellite images without requiring actual high-resolution aerial images for training
Solution Approach 2:
The patent performs preliminary actions by pre-training the MLN on synthesized data before deploying it for actual 3D reconstruction. The training phase prepares the network in advance to handle the resolution mismatch between satellite images and ground truth data, allowing efficient processing of large areas while maintaining accuracy
2Measurement precision
If aerial images are used for training, then higher resolution ground truth data can be obtained, but capturing time increases significantly
Solution Approach 1:
Instead of capturing actual aerial images, the patent copies the essential geometric and photometric properties by rendering synthetic images and depth maps from existing 3D surface models. This approach provides unlimited high-resolution training data without any field campaigns or time-consuming aerial photography
Solution Approach 2:
The patent replaces the mechanical system of aerial image capture with a computational rendering system. Rather than physically flying aircraft to capture ground truth data, the system uses computer graphics rendering to generate synthetic training pairs, eliminating the time and resource constraints of aerial photography
3Device complexity
If traditional 3D reconstruction methods are used, then processing is simpler, but accuracy and efficiency for large areas are reduced
Solution Approach 1:
The patent replaces traditional geometric processing methods with a deep learning-based approach. The MLN automatically learns complex relationships between satellite images and depth maps, handling the complexity of large-area reconstruction while maintaining simplicity in deployment through a unified neural network model that processes entire scenes efficiently
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present disclosure relates to a method for 3D reconstruction from satellite imagery using deep learning, said method comprising providing (101) at least two overlapping 2D satellite images, providing (102) imaging device parameters for the at least two overlapping 2D satellite images, providing (103) at least one trained Machine Learning Network, MLN, able to predict depth maps, said trained MLN being trained on a training set comprising multi-view geocoded 3D ground truth data and predicting (104) a depth map of the at provided at least two 2D satellite images using the trained at least one MLN and based on the corresponding imaging device parameters.