Object Detector Feature Maps for CNN Stability Evaluation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for evaluating the stability of Convolutional Neural Networks (CNNs) in autonomous driving are limited and difficult to apply to general cases, lacking a comprehensive way to determine if the CNN is operating stably, which is crucial for ensuring safe operation of autonomous vehicles.

Innovation Solution

A method for evaluating object detection processes using representative feature maps of object detectors, involving the generation and comparison of feature maps to determine reliability scores and provide logical grounds for autonomous driving decisions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If conventional Class Activation Map configuration is used to evaluate CNN stability, then evaluation can be performed for specific configured cases, but the method cannot be applied to general CNN configurations

Engineering Contradiction:
Improveapplicability to general CNN casesVSAvoidconfiguration requirements
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent creates a universal evaluation method that works across different CNN configurations by generating representative feature maps that capture essential detection patterns without requiring specific Class Activation Map configurations. The method uses multiple diverse test images to create comprehensive feature map representations that are applicable to general object detection scenarios.

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

2Reliability

If comprehensive evaluation of CNN stability is implemented, then reliable detection can be ensured, but the evaluation process becomes complex and difficult to apply

Engineering Contradiction:
ImproveCNN operation stabilityVSAvoidevaluation process complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent creates simplified copies of CNN internal states by generating representative feature maps that replicate essential detection information. These feature maps serve as tractable representations of the complex CNN processing, allowing evaluation without directly analyzing the full complexity of the neural network operations.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The method performs partial evaluation by focusing on representative feature maps from key layers rather than comprehensively analyzing all CNN operations. This selective approach provides sufficient reliability for stability assessment while keeping the evaluation process manageable.

Inventive Principle:
Principle #16Partial or excessive action

3Measurement precision

If detailed feature map analysis is performed for each object detection, then detection reliability can be assessed, but processing time and computational resources increase

Engineering Contradiction:
Improvedetection evaluation accuracyVSAvoidevaluation processing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent extracts only the essential information needed for evaluation by generating representative feature maps that capture key detection patterns. Instead of analyzing all feature maps in detail, the method selects and analyzes representative samples that provide sufficient evaluation accuracy with reduced computational overhead.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP3690733B1Method and device for supporting administrators to evaluate object detecting processes of object detectors to provide logical grounds of autonomous driving
Publication Date: 2026.01.14 STRADVISION
  • EP3690733B1 patent drawingFigure 1
  • EP3690733B1 patent drawingFigure 2
  • EP3690733B1 patent drawingFigure 3

AI summary

A method for supporting at least one administrator to evaluate detecting processes of object detectors to provide logical grounds of an autonomous driving is provided. And the method includes steps of: (a) a computing device instructing convolutional layers, included in an object detecting CNN which has been trained before, to generate reference convolutional feature maps by applying convolutional operations to reference images inputted thereto, and instructing ROI pooling layers included therein to generate reference ROI-Pooled feature maps by pooling at least part of values corresponding to ROIs on the reference convolutional feature maps; and (b) the computing device instructing a representative selection unit to classify the reference ROI-Pooled feature maps by referring to information on classes of objects included in their corresponding ROIs on the reference images, and to generate at least one representative feature map per each class.