Pixel-Level Coordinate Regression With Sparse Cross-View Annotations

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing computer vision systems struggle to accurately regress 3D scene coordinates in unseen environments due to their reliance on scene-specific training, requiring complex pipelines and sparse feature matching, which limits their scalability and accuracy.

Innovation Solution

A machine learning model is trained using sparse pixel-level annotations from similar images, leveraging an encoder-decoder architecture with cross-view completion pre-training, and a feature mixer to generate dense pixel-level predictions, enabling scene-agnostic regression.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If structure-based visual localization frameworks use sparse feature matching to estimate pose, then pose estimation capability is improved, but device complexity increases due to multi-step pipeline and storage requirements for 3D points and key point descriptors

Engineering Contradiction:
Improvepose estimation accuracyVSAvoidlocalization pipeline complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts only the essential 3D coordinate information needed for pose estimation, discarding the complex multi-step pipeline of traditional structure-based methods. By directly regressing 3D coordinates from image pixels using a neural network, the system eliminates the need for sparse feature matching, key point descriptor storage, and RANSAC-based pose solving, thereby reducing device complexity while maintaining pose estimation capability

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent replaces the mechanical multi-step localization pipeline (feature detection, descriptor matching, 2D-3D correspondence establishment, RANSAC pose solving) with a direct neural network-based regression system. This substitution transforms a complex sequential processing system into a unified end-to-end learning model, significantly simplifying the overall system architecture

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Measurement precision

If coordinate regression approaches embed 3D scene coordinates directly into model parameters, then regression accuracy is improved, but adaptability deteriorates as the approach is limited to specific scenes

Engineering Contradiction:
Improvecoordinate regression accuracyVSAvoidscene generalization capability
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent creates a universal coordinate regression model that can handle multiple scenes through a single unified architecture. By using a scene-agnostic neural network that learns to map image pixels to 3D coordinates without embedding scene-specific parameters, the system achieves multi-functionality across different environments. The model processes queries from any scene by leveraging learned geometric priors and relationships, eliminating the need for scene-specific model instances

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent changes the parameter representation from embedded 3D scene coordinates (scene-specific) to learned feature representations that capture geometric relationships (scene-agnostic). By transforming the model parameters from explicit coordinate embeddings to implicit geometric priors learned through training on diverse scenes, the system achieves both accuracy and adaptability across unseen environments

Inventive Principle:
Principle #35Parameter changes

3Manufacturing precision

If dense annotations are used for training coordinate regression models, then manufacturing precision is improved, but loss of substance increases due to the impracticality of obtaining dense annotations for unknown test scenes

Engineering Contradiction:
Improveannotation precisionVSAvoidannotation availability
Core Design Contradiction:
Manufacturing precisionVSLoss of substance

Solution Approach 1:

The patent adopts partial action by using sparse annotations (only key 3D points) instead of requiring dense annotations for all pixels. The neural network learns to generalize from these sparse labeled points to predict coordinates for all pixels in the scene. This approach makes training practical for unknown test scenes where obtaining dense annotations would be prohibitively expensive or impossible, while still achieving high precision through the model's generalization capability

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4471663B1Scene-agnostic regression on pixel-level annotations
Publication Date: 2026.05.20 NAVER CORP
  • EP4471663B1 patent drawingFigure 1
  • EP4471663B1 patent drawingFigure 2
  • EP4471663B1 patent drawingFigure 3

AI summary

A computer-implemented method of training a machine learning model for regression on pixel-level annotations in images is disclosed. The method comprises pre-training an image encoder and a decoder for cross-view completion, constructing training tuples, each comprising a first image, associated with dense pixel-level annotations, and one or more second images, each associated with sparse pixel-level annotations. The image encoder generates first image tokens and second image tokens from the first and second images. A feature mixer generates sets of augmented second image tokens by augmenting the second image tokens with the associated sparse pixel-level annotations. The method further comprises processing, by the decoder, the first image tokens and the augmented second image tokens to generate prediction data for the first image, the prediction data comprising dense pixel-level predictions, and fine-tuning the machine learning model based on the prediction data and the dense pixel-level annotations.