Cross-View Discriminator Training With Progressive Negative Examples
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for ground-to-aerial cross-view matching do not account for the order of aerial-view images during training, leading to suboptimal accuracy in discriminator performance.
Innovation Solution
A training apparatus and method that involves two-phase training of a discriminator, using positive and negative examples of aerial-view images with varying degrees of similarity to ground-view images, employing curriculum learning to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If arbitrary order of aerial-view images is used in training, then training process is simple, but matching accuracy deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-sorting negative example aerial-view images based on their scenery type similarity to the ground-view image before training. This preliminary organization of training data in ascending order of difficulty (from most different to most similar scenery types) enables the discriminator to progressively learn distinguishing features, thereby improving matching accuracy while maintaining a systematic and manageable training process
2Loss of time
If single-phase training is used, then training time is short, but discriminator accuracy deteriorates
Solution Approach 1:
The patent applies segmentation by dividing the training process into two distinct phases: first-phase training using only first level negative examples (scenery of different type), and second-phase training using second level negative examples (scenery of same type). This segmentation allows the discriminator to progressively master easier distinctions before tackling more challenging ones, improving overall accuracy while maintaining efficient training through structured progression
3Reliability
If discriminator is trained with all negative examples simultaneously, then training completeness is high, but training efficiency deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-classifying and organizing negative examples into two levels based on scenery type similarity before training begins. This preliminary classification enables the discriminator to first learn from clearly distinguishable examples (first level), then progressively learn from more challenging examples (second level), thereby maintaining training completeness while significantly improving training efficiency through graduated difficulty progression
Data Source
AI summary
The training apparatus (2000) performs a first phase training and a second phase training of a discriminator (10). The discriminator (10) acquires a ground-view image and an aerial-view image, and determines whether the acquired ground-view image matches the acquired aerial-view image. The first phase training is performed using a ground-view image and a first level negative example of aerial-view image. The first level negative example of aerial-view image includes scenery of a different type from scenery in the ground-view image. The second phase training is performed using the ground-view image and a second level negative example of aerial-view image. The second level negative example of aerial-view image includes scenery of a same type as scenery in the ground-view image.