Image Correspondence Learning With Pyramidal Matching Priors

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing learning-based methods for extracting local image descriptors require annotated training data sets with pixel-level correspondences, limiting the potential of available image pairs and restricting the quality of training results.

Innovation Solution

A method for unsupervised learning of local image descriptors using a pyramidal structure to enforce local consistency and uniqueness priors, allowing the neural network to learn optimal descriptors without supervision, enabling the identification of corresponding portions in images with differing parameters.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If supervised learning approaches are used to extract local image descriptors, then the quality of identified correspondences is improved, but the requirement for large amounts of annotated training data increases

Engineering Contradiction:

Improvequality of correspondencesVSAvoidamount of training data

Core Design Contradiction:

Measurement precisionVSQuantity of substance

Solution Approach 1:

The system performs self-supervised learning by automatically generating training labels from the image data itself through consistency regularization across augmented views, eliminating the need for external annotated training data while maintaining high correspondence quality

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The method pre-computes image augmentations and their transformations to generate pseudo-labels before the actual learning process, allowing the model to be trained on unlabeled data with pre-generated supervision signals

Inventive Principle:

Principle #10Preliminary action

2Measurement precision

If ground-truth correspondences are required for training, then the accuracy of local image descriptors is improved, but the limitation of available training data increases

Engineering Contradiction:

Improveaccuracy of descriptorsVSAvoidexploitation of image pairs

Core Design Contradiction:

Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The system generates its own training labels by computing consistency relationships between augmented views of images, allowing any image pair to be used for training without requiring pre-existing ground-truth correspondences

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The training framework can process any image pair regardless of whether ground-truth correspondences are available, making the approach universally applicable to both labeled and unlabeled data for descriptor learning

Inventive Principle:

Principle #6Universality (Multi-functionality)

Data Source

PatentUS12626489B2System and method using pyramidal and uniqueness matching priors for identifying correspondences between images

Publication Date: 2026.05.12 NAVER CORP

AI summary

A computer-implemented method includes: obtaining a pair of images depicting a same scene, the pair of images including a first image with a first pixel grid and a second image with a second pixel grid, the first pixel grid different than the second pixel grid; by a neural network module having a first set of parameters: generating a first feature map based on the first image; and generating a second feature map based on the second image; determining a first correlation volume based on the first and second feature maps; iteratively determining a second correlation volume based on the first correlation volume; determining a loss for the first and second feature maps based on the second correlation volume; generating a second set of the parameters based on minimizing a loss function using the loss; and updating the neural network module to include the second set of parameters.