Instance-Aware Sampling for Rare-Class Object Detection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing methods for addressing class imbalance in object detection, such as Repeat Factor Sampling (RFS), fail to adequately represent rare classes due to imbalances between the number of images and bounding boxes, leading to biased training and underrepresentation of minor classes.

Innovation Solution

Instance-aware Repeat Factor Sampling (I-RFS) method that considers both the number of images and bounding boxes for each class, using combined factors to determine the sampling strategy, ensuring rare classes are prioritized.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If class-balanced sampling is used to up-sample minor classes, then performance on minor classes is improved, but training becomes biased towards augmented minor classes and available data representation of major classes is not properly exploited

Engineering Contradiction:

Improveperformance on minor classesVSAvoidtraining bias

Core Design Contradiction:

ReliabilityVSEase of operation

Solution Approach 1:

The patent changes the sampling parameters by introducing instance-aware repeat factors that are calculated based on the actual instance counts of each class. This dynamically adjusts the sampling strategy to balance between up-sampling minor classes and maintaining proper representation of major classes, resolving the bias issue while improving minor class performance

Inventive Principle:

Principle #35Parameter changes

2Reliability

If larger up-sampling and down-sampling ratio is used to handle significant imbalance, then minor classes are better represented, but major classes are under-represented and training becomes biased

Engineering Contradiction:

Improverepresentation of minor classesVSAvoidrepresentation of major classes

Core Design Contradiction:

ReliabilityVSQuantity of substance

Solution Approach 1:

The patent applies local quality by calculating separate repeat factors for each class based on their individual instance counts and image distributions. This allows tailored sampling rates for different classes - higher for minor classes and lower for major classes - rather than applying a uniform sampling ratio, thus preventing under-representation of major classes

Inventive Principle:

Principle #3Local quality

3Productivity

If single-phase training with modified loss function is used, then optimization is carried out on long-tailed data, but minor classes are significantly under-represented due to large imbalance

Engineering Contradiction:

Improvetraining efficiencyVSAvoidperformance on minor classes

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The patent performs preliminary action by calculating instance-aware repeat factors before training and using these factors to reweight the training samples. This pre-processing step ensures that minor classes are adequately represented in the training set before the optimization process begins, preventing their under-representation during single-phase training

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS12482238B2Instance-aware repeat factor sampling method for imbalanced data in object detection

Publication Date: 2025.11.25 ROBERT BOSCH GMBH

AI summary

An approach for handling long-tail distribution with data imbalance. Disclosed embodiments improve Repeat Factor Sampling methods by considering both images and bounding boxes (i.e., instances) to generate improved sets of training data. Disclosed embodiments may be useful in image analysis domains, such as object detection and classification.