Balanced Training Data Collage for Class-Imbalanced Object Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The challenges of generating high-quality pseudo-labels, addressing class imbalance, and mitigating noise in labeled and pseudo-labeled data hinder the effectiveness of semi-supervised learning in object detection, particularly in complex environments like autonomous driving, leading to biased models and suboptimal performance.

Innovation Solution

A method for generating training data by cutting out objects from sample images based on their frequency of occurrence, creating collages of underrepresented objects, and adding them to the training dataset to balance class frequencies, thereby reducing noise and improving model generalization.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If semi-supervised learning uses teacher model to generate pseudo-labels for unlabeled data, then the dependence on large labeled datasets is reduced, but the quality of pseudo-labels deteriorates leading to noisy or incorrect labels that mislead the student model

Engineering Contradiction:
Improveamount of labeled dataVSAvoidquality of pseudo-labels
Core Design Contradiction:
Quantity of substanceVSReliability

Solution Approach 1:

The patent introduces an intermediary verification mechanism where a second teacher model or verification module acts as a mediator to validate pseudo-labels before they are used to train the student model. This intermediary layer filters out low-quality pseudo-labels and prevents noisy labels from misleading the student model, thus maintaining reliability while still reducing dependence on large labeled datasets.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent implements a feedback mechanism where the student model's predictions are compared with pseudo-labels, and the discrepancies are used to refine the teacher model's future pseudo-label generation. This feedback loop continuously improves pseudo-label quality over time, allowing the system to maintain high reliability even with reduced labeled data while adapting to the specific characteristics of the unlabeled dataset.

Inventive Principle:
Principle #23Feedback

2Ease of manufacture

If training data contains underrepresented object classes, then the model can be trained on available data, but class imbalance causes the model to become biased towards majority classes and fail to accurately detect underrepresented classes

Engineering Contradiction:
Improveease of data collectionVSAvoiddetection accuracy of underrepresented classes
Core Design Contradiction:
Ease of manufactureVSReliability

Solution Approach 1:

The patent applies preliminary data augmentation and re-sampling techniques during the data preparation phase to balance class distributions before training begins. By pre-processing the training data to ensure adequate representation of underrepresented classes through techniques like oversampling, undersampling, or synthetic data generation, the model is exposed to balanced class frequencies from the start, preventing bias towards majority classes while avoiding the need for difficult data collection.

Inventive Principle:
Principle #10Preliminary action

3Productivity

If training data contains noisy or missing annotations, then the dataset can be used as-is, but the noise in labels complicates the training process and leads to suboptimal model performance

Engineering Contradiction:
Improvetraining speedVSAvoidmodel performance
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent applies partial cleaning by selectively processing only the most problematic noisy samples or focusing cleaning efforts on critical object classes, rather than exhaustively cleaning every label. This partial action approach maintains training speed by avoiding complete re-annotation while still removing the most harmful noise that would significantly degrade model performance, achieving a practical balance between productivity and reliability.

Inventive Principle:
Principle #16Partial or excessive action

4Ease of manufacture

If the training dataset is imbalanced with majority and minority classes, then training can proceed with available data, but the model becomes biased towards majority class and overfits to common patterns

Engineering Contradiction:
Improveease of dataset preparationVSAvoidclass distribution balance
Core Design Contradiction:
Ease of manufactureVSStability of the object's composition

Solution Approach 1:

The patent dynamically adjusts training parameters such as loss function weights, learning rates, and sampling probabilities based on the observed class distribution in the training data. By automatically modifying these parameters to compensate for class imbalance, the system maintains stable and balanced class representation during training without requiring manual dataset re-balancing, thus preserving ease of preparation while achieving distribution stability.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4712044A1A method, apparatus, and computer-readable medium for generating balanced datasets for machine learning
Publication Date: 2026.03.18 CONTINENTAL AUTOMOTIVE TECHNOLOGIES GMBH
  • EP4712044A1 patent drawingFigure 1
  • EP4712044A1 patent drawingFigure 2
  • EP4712044A1 patent drawingFigure 3

AI summary

Embodiments of the present disclosure relate to a method for generating training data to be used in training a machine learning model for a computer vision task. A training method and data 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 generating training data to be used in training a machine learning model comprises obtaining information on a frequency of occurrence of object classes in the training dataset, cutting out, based on the frequency of occurrence, objects from sample images obtained from the training dataset, and generating training data comprising a collage of at least two of the cut-out objects.