Neural Radiance Fields for Automated Object Detection Training Data

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Collecting large amounts of manually labelled training data for object detection in autonomous systems is time-consuming, labour-intensive, and costly, particularly in applications like autonomous driving and industrial automation.

Innovation Solution

Utilize a first neural radiance field (NeRF) to generate feature values for optical images and a second NeRF to encode sensor data from multiple modalities, allowing for efficient generation of labelled training data without manual labelling, using techniques like CLIP features for text-driven 3D labelling.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual labelling of training data is used, then ground truth information can be obtained, but time consumption and labour intensity increase significantly

Engineering Contradiction:
Improveground truth information qualityVSAvoiddata collection time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent uses a source object detector (trained on manually labelled data) to automatically generate labels for target sensor data. The source detector's object detection results are copied and adapted to create ground truth information for the target domain, eliminating the need for manual labelling of the target dataset while maintaining label quality.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent introduces a domain adaptation module as an intermediary between the source object detector and the target sensor data. This intermediary transforms and adapts the source detector's predictions to match the target domain characteristics, enabling automatic label generation without direct manual annotation of target data.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If manually labelled training data is collected, then accurate ground truth information is obtained, but cost increases significantly

Engineering Contradiction:
Improveground truth information qualityVSAvoiddata generation cost
Core Design Contradiction:
Measurement precisionVSEase of manufacture

Solution Approach 1:

The patent copies the labelling capability from a pre-trained source object detector to generate labels for target data. This approach replaces expensive manual labelling processes with automated detection, significantly reducing the cost of producing ground truth information while maintaining accuracy.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The system enables self-service labelling where the source object detector automatically generates its own training data for the target domain. The detector performs both detection and self-labeling functions, eliminating the need for external human annotators and reducing operational costs.

Inventive Principle:
Principle #25Self-service

3Adaptability or versatility

If domain adaptation is performed, then sensor setup variations are handled, but computational complexity increases

Engineering Contradiction:
Improvesensor setup flexibilityVSAvoiddomain adaptation complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent performs domain adaptation in advance by using the source object detector to generate adapted labels before actual target data processing. The adaptation process is executed beforehand to create a pre-adapted model or label set, reducing the computational burden during real-time or production operations.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent copies the domain adaptation logic from the source detector framework and applies it to the target domain. By reusing and adapting existing detection algorithms rather than developing new ones, the system manages computational complexity while achieving sensor setup flexibility.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4604075A1Device and method for generating training data for an object detector
Publication Date: 2025.08.20 ROBERT BOSCH GMBH
  • EP4604075A1 patent drawingFigure 1
  • EP4604075A1 patent drawingFigure 2~3
  • EP4604075A1 patent drawingFigure 4

AI summary

According to various embodiments, a method for generating training data for an object detector is described, comprising receiving a plurality of optical images of a scene, each camera showing the scene from a respective viewing direction of a plurality of different viewing directions, receiving a plurality of sensor data elements, each sensor data element including sensor data other than optical image data of the scene from a respective sensing direction of a plurality of different sensing directions, training a first neural radiance field using the plurality of optical images to generate, for each 3D point of the scene, a respective value of a predetermined feature, training a second neural radiance field using the plurality of sensor data elements to generate, for each 3D point of the scene, a respective sensor data value and generating training data elements for the object detector using the first and the second neural radiance field.