Ground-to-Aerial Image Augmentation with Temporary-Feature Removal
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems for ground-to-aerial cross-view matching are prone to overfitting due to random region erasing techniques that do not effectively address temporary features in training images.
Innovation Solution
An image augmentation apparatus and method that systematically removes specific types of objects from ground-view and aerial-view images to generate an augmented training dataset, enhancing robustness to temporary features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If random region erasing is used to prevent overfitting, then the model becomes more robust to occlusion, but the erasing only removes partial regions and does not effectively address temporary features
Solution Approach 1:
The patent applies local quality by selectively removing specific types of objects (temporary features) rather than using random erasing. The system identifies and removes regions containing temporary features such as construction vehicles, temporary structures, or seasonal objects, while preserving permanent features. This targeted approach allows the model to learn robust representations by focusing on stable, meaningful elements in the image.
Solution Approach 2:
Instead of randomly erasing regions and hoping to remove temporary features, the patent inverts the approach by actively identifying and removing specific temporary feature regions. The system uses object detection or segmentation to locate temporary features and then removes them, rather than relying on random erasing to coincidentally remove these features.
2Reliability
If random rectangle regions are erased from training images, then overfitting is reduced, but the training images do not effectively simulate real-world occlusion scenarios
Solution Approach 1:
The patent removes specific temporary features from images to create training data that better represents real-world scenarios. By targeting and removing temporary features (such as construction equipment, temporary structures, or seasonal objects), the system creates more realistic occlusion patterns that reflect actual environmental conditions, rather than using arbitrary random erasing.
Solution Approach 2:
The system performs preliminary identification and removal of temporary features before training the discriminator. This preliminary action ensures that the training images are pre-conditioned to exclude temporary features, allowing the model to learn from more representative and realistic image pairs from the beginning of training.
3Ease of manufacture
If temporary features are not removed from training images, then the training process remains simple, but the discriminator learns to rely on transient information leading to overfitting
Solution Approach 1:
The patent performs preliminary removal of temporary features from training images before the discriminator training begins. This pre-processing step ensures that the training data is cleaned of transient information that could cause overfitting, allowing the discriminator to learn from stable, meaningful features from the start of training.
Solution Approach 2:
The system extracts and removes temporary features from training images, separating them from permanent features. By taking out temporary elements (such as construction vehicles, temporary structures, or seasonal objects), the system creates training data that focuses on stable, generalizable features, preventing the discriminator from learning to rely on transient information.
Data Source
AI summary
An image augmentation apparatus (2000) performs an image augmentation process on an original training dataset (40) to generate an augmented training dataset (50). The original training dataset (40) includes an original ground-view image (42) and an original aerial-view image (44). The image augmentation process includes a removal process in which partial regions including objects of a specific type are removed. The image augmentation apparatus (2000) performs the image augmentation process on the original ground-view image (42) to generate an augmented ground-view image (52), on the original aerial-view image (44) to generate an augmented aerial-view image (54), or both.