Neural Network Blind Spot Implementation for Object Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current computer vision models lack the ability to implement blind spots, leading to unnecessary object detections and excessive resource consumption, as they fail to selectively identify objects of interest, resulting in inefficient notifications and poor user experience.

Innovation Solution

The implementation of a neural network model that introduces blind spots by training on datasets with intentionally unannotated objects of interest, allowing the model to learn and classify objects based on their position within the field of view, thereby reducing unnecessary detections and resource consumption.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If computer vision models detect all objects of interest within the field of view, then detection completeness is improved, but resource consumption and notification overload worsen

Engineering Contradiction:
Improvedetection completenessVSAvoidresource consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The field of view is segmented into zones of interest and blind spots. The neural network is trained to selectively detect objects only in zones of interest while ignoring objects in blind spots. This segmentation allows the system to maintain detection completeness for important areas while reducing resource consumption by ignoring less important areas.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Different regions of the field of view are assigned different detection qualities. Zones of interest receive high detection sensitivity while blind spots receive no detection. This local quality differentiation enables the system to optimize resource allocation by concentrating detection efforts where they are most needed.

Inventive Principle:
Principle #3Local quality

2Reliability

If computer vision models detect all objects of interest within the field of view, then detection completeness is improved, but notification overload and human attention requirements worsen

Engineering Contradiction:
Improvedetection completenessVSAvoidnotification efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The detection system is segmented to generate notifications only for objects detected in zones of interest. By excluding blind spots from detection, the system reduces the total number of notifications while maintaining completeness for important areas, thereby improving notification efficiency and reducing human attention requirements.

Inventive Principle:
Principle #1Segmentation

3Use of energy by moving object

If computer vision models implement selective detection with blind spots, then resource consumption is reduced, but detection completeness worsens

Engineering Contradiction:
Improveresource consumptionVSAvoiddetection completeness
Core Design Contradiction:
Use of energy by moving objectVSReliability

Solution Approach 1:

The system applies different detection qualities to different spatial regions. Zones of interest maintain high detection sensitivity and completeness, while blind spots are intentionally excluded. This local quality approach ensures that resource reduction does not compromise overall detection completeness, as the excluded areas are deliberately chosen to be less critical.

Inventive Principle:
Principle #3Local quality

4Adaptability or versatility

If neural networks are trained with intentionally unannotated objects, then blind spot capability is improved, but training complexity worsens

Engineering Contradiction:
Improveblind spot capabilityVSAvoidtraining complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

Instead of training the neural network to detect all objects and then filtering results, the training approach is inverted: objects in blind spot regions are intentionally left unannotated during training. This causes the network to learn not to detect objects in those regions, directly embedding blind spot capability into the model without requiring post-processing filtering.

Inventive Principle:
Principle #13The other way round (Inversion)

Data Source

PatentEP3696719B1Blind spot addition to neural networks
Publication Date: 2024.12.04 SAP SE
  • EP3696719B1 patent drawingFigure 1
  • EP3696719B1 patent drawingFigure 2
  • EP3696719B1 patent drawingFigure 3

AI summary

Techniques for implementing blind spots into neural network models are disclosed. In some example embodiments, a computer-implemented method comprises obtaining an image captured within a field of view of an image capture device and comprising an object of a particular type occupying a particular position within the field of view, and determining a confidence value for the object based on the particular position of the object using a neural network model. The confidence value represents a likelihood that the object is an object of interest, and the neural network model is trained to generate lower confidence values for objects of the particular type when the objects of the particular type are occupying the particular position than when the objects of the particular type are not occupying the particular position.