Ground Truth Label Correction for Noisy Vision Training Data

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The quality of pseudo-labels in semi-supervised learning is often hampered by noise, class imbalance, and inherent noise in labeled and pseudo-labeled data, leading to suboptimal model performance, particularly in complex environments like autonomous driving.

Innovation Solution

A method that modifies sample images using augmentation and obtains predicted labels during inference-time to correct erroneous, missing, or noisy ground truth labels, and generates a more balanced training dataset by augmenting underrepresented classes, thereby improving the reliability and accuracy of training data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If semi-supervised learning uses pseudo-labels for unlabeled data, then the model can leverage more data, but the quality of pseudo-labels contains noise and errors that degrade performance

Engineering Contradiction:
Improveamount of training dataVSAvoidquality of labels
Core Design Contradiction:
Quantity of substanceVSReliability

Solution Approach 1:

The patent introduces an intermediary verification process where a teacher model generates pseudo-labels that are then validated against multiple criteria (confidence thresholds, consistency checks, cross-validation) before being used to train the student model. This intermediary step filters out noisy labels while preserving useful unlabeled data.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent implements feedback mechanisms where the student model's predictions are compared against teacher model predictions and ground truth labels to identify and correct erroneous pseudo-labels. The system continuously refines label quality through iterative feedback loops during training.

Inventive Principle:
Principle #23Feedback

2Productivity

If the training dataset has class imbalance with underrepresented classes, then the model can be trained faster, but the model fails to generalize well across all object categories

Engineering Contradiction:
Improvetraining speedVSAvoidgeneralization capability
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent applies local quality enhancement by identifying underrepresented classes and applying targeted data augmentation and oversampling specifically to those classes. The system adjusts the training strategy locally for minority classes while maintaining efficient processing for majority classes.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent performs preliminary analysis of class distribution before training and proactively balances the dataset through augmentation and resampling. This preliminary action ensures all classes are adequately represented before the main training process begins, preventing generalization issues.

Inventive Principle:
Principle #10Preliminary action

3Productivity

If ground truth labels contain noise or missing annotations, then data collection is faster and cheaper, but the training process produces suboptimal model performance

Engineering Contradiction:
Improvedata collection efficiencyVSAvoidlabel accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent uses feedback loops where model predictions are compared against ground truth labels to identify inconsistent or erroneous annotations. The system detects noise patterns and missing annotations through confidence scoring and consistency checks, then corrects or flags these labels for re-verification.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent introduces an intermediary validation layer that assesses ground truth labels before they are used for training. The teacher model and verification mechanisms act as intermediaries to filter out noisy labels while preserving the efficiency of using initially collected data.

Inventive Principle:
Principle #24Intermediary (Mediator)

4Adaptability or versatility

If image augmentation is applied to generate diverse training samples, then the model becomes more robust to variations, but the computational complexity increases

Engineering Contradiction:
Improverobustness to variationsVSAvoidcomputational complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent applies partial augmentation by selecting and applying only the most effective augmentation techniques for specific scenarios and classes. The system dynamically adjusts the level and type of augmentation based on data characteristics, applying stronger augmentation to underrepresented classes and weaker augmentation to well-represented classes.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The patent implements dynamic augmentation strategies where the type and intensity of image transformations are adjusted during training based on model performance and data characteristics. The system adapts the augmentation pipeline in real-time to balance robustness gains with computational efficiency.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4711990A1A method, apparatus, and computer-readable medium for correcting ground truth labels in training datasets
Publication Date: 2026.03.18 CONTINENTAL AUTOMOTIVE TECHNOLOGIES GMBH
  • EP4711990A1 patent drawingFigure 1
  • EP4711990A1 patent drawingFigure 2
  • EP4711990A1 patent drawingFigure 3

AI summary

Embodiments of the present disclosure relate to a method for correcting ground truth labels used in training a machine learning model for a computer vision task. A training method and ground truth labels used for training a machine learning model are also disclosed. A machine learning model, a computer program, a computer-readable data carrier and an apparatus are also presented. The method for correcting ground truth labels comprises obtaining sample images and ground truth labels for the sample images, training the machine learning model using the sample images and the ground truth labels, modifying the sample images, obtaining, by the machine learning model, predicted labels for the modified sample images, and correcting the ground truth labels based on a comparison of the ground truth labels with the predicted labels.