3D Object Detection Using Geometric Constraints With Less 3D Annotation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The need for large amounts of 3D-annotated training data poses a significant challenge in 3D object detection, as annotating 3D images with accurate 3D bounding boxes is a burdensome task for humans, while existing methods require extensive manual labor.

Innovation Solution

The method exploits expected geometric relationships between predicted 3D boundary objects to reduce the amount of 3D annotation required, utilizing 2D annotation data and incorporating geometric constraints into the cost function during training, allowing 3D object detection with state-of-the-art accuracy using fewer 3D-annotated inputs.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If 2D object detection algorithms are used, then the system is simple and fast, but it cannot detect objects in 3D space or determine depth information

Engineering Contradiction:
Improvedepth detection capabilityVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies dimensionality change by transforming 2D image data into 3D spatial understanding through depth map generation. The neural network processes 2D images and outputs depth information, effectively adding a third dimension (depth) to the traditional 2D detection framework, enabling 3D object detection without requiring complex multi-camera setups

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If multiple cameras are used to capture images from different angles, then 3D object detection capability is improved, but the system complexity and cost increase

Engineering Contradiction:
Improve3D detection capabilityVSAvoidnumber of cameras
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent uses a single camera to capture multiple views of the scene by moving the camera to different positions. This copying approach allows the system to obtain multi-angle images without requiring multiple simultaneous cameras, reducing hardware complexity while maintaining 3D detection capability through sequential image capture and processing

Inventive Principle:
Principle #26Copying

3Device complexity

If a single camera captures images from different positions, then system complexity is reduced, but the time to capture all necessary views increases

Engineering Contradiction:
Improvenumber of camerasVSAvoidimage capture time
Core Design Contradiction:
Device complexityVSLoss of time

Solution Approach 1:

The patent pre-defines multiple capture positions and uses a positioning mechanism to automatically move the camera to each predetermined position in sequence. This preliminary setup of capture positions allows the system to efficiently capture all necessary views without manual intervention, reducing the overall time required compared to arbitrary positioning while using only a single camera

Inventive Principle:
Principle #10Preliminary action

4Measurement precision

If complex neural network architectures are used to improve detection accuracy, then detection precision improves, but computational resources and processing time increase

Engineering Contradiction:
Improveobject detection accuracyVSAvoidcomputational resource consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent segments the object detection task into multiple processing stages: depth map generation from 2D images, 3D bounding box parameter extraction, and classification. This segmentation allows the neural network to process information in manageable steps, improving accuracy for each specific task while optimizing computational resource usage compared to attempting all tasks simultaneously in a single complex model

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4055518B13D object detection
Publication Date: 2026.04.29 FIVE AI LTD
  • EP4055518B1 patent drawingFigure 1
  • EP4055518B1 patent drawingFigure 2
  • EP4055518B1 patent drawingFigure 3

AI summary

A method of training a 3D structure detector to detect 3D structure in 3D structure representation, the method comprising the following steps: receiving, at a trainable 3D structure detector, a set of training inputs, each training input comprising at least one 3D structure representation; the 3D structure detector determining, for each training input, a set of predicted 3D objects for the at least one 3D structure representation of that training input; and training the 3D structure detector to optimize a cost function, wherein the cost function penalizes deviation from an expected geometric relationship between the set of predicted 3D objects determined for each training input.