Camera-Conditioned Neural Network Training for Metric 3D Scene Depth

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing Monocular Depth Estimation (MDE) methods require training and testing on datasets with similar camera intrinsic parameters and limited scene diversity, leading to poor generalization in real-world scenarios with arbitrary camera setups and uncontrolled environments.

Innovation Solution

A neural network training method that determines a 3D representation of a scene from a single image by estimating camera intrinsic parameters and using a camera-conditioned depth feature map, minimizing losses to predict depth in meters without additional external information, and jointly training multiple networks for improved robustness and consistency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If existing MDE methods are trained on datasets with similar camera intrinsic parameters and limited scene diversity, then training accuracy is improved, but generalization performance to real-world scenarios deteriorates

Engineering Contradiction:
Improvetraining accuracyVSAvoidgeneralization performance
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent transforms the depth prediction task from predicting depth in arbitrary units to predicting depth in meters by incorporating camera intrinsic parameters as conditional inputs. This parameter transformation enables the model to generalize across different camera setups while maintaining accurate depth estimation, resolving the contradiction between training accuracy and generalization performance.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent separates camera intrinsic parameters from scene features by using camera-conditioned depth feature maps. The camera parameters are processed independently and then integrated with scene features through cross-attention mechanisms, allowing the model to handle arbitrary camera configurations while maintaining robust training performance.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If camera intrinsic parameters are incorporated into the neural network for accurate depth prediction, then depth accuracy in meters is improved, but model complexity increases

Engineering Contradiction:
Improvedepth accuracyVSAvoidmodel complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent introduces camera-conditioned depth feature maps as intermediaries between camera intrinsic parameters and the depth prediction output. These feature maps serve as a bridge that integrates camera parameters with scene features through cross-attention mechanisms, achieving accurate metric depth prediction while managing model complexity through structured information flow.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent adds a conditional dimension to the depth prediction by incorporating camera intrinsic parameters as additional inputs. This transforms the prediction from a 2D depth map to a 3D representation that includes camera parameter conditioning, enabling accurate metric depth estimation without proportionally increasing overall model complexity.

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

3Reliability

If multiple neural networks are jointly trained with consistency loss, then robustness and consistency across different cameras are improved, but training time and computational resources increase

Engineering Contradiction:
ImproverobustnessVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent merges multiple neural networks into a unified joint training framework where networks share common structures and parameters. By combining multiple networks and applying consistency loss across them, the system achieves improved robustness and cross-camera consistency while optimizing training efficiency through shared computational components.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP4614441A1Method and system for training a neural network for generating a three-dimensional representation of a scene
Publication Date: 2025.09.10 TOYOTA JIDOSHA KK
  • EP4614441A1 patent drawingFigure 1~2
  • EP4614441A1 patent drawingFigure 3~4
  • EP4614441A1 patent drawingFigure 5

AI summary

The invention relates to a method and a system for training a neural network for generating a 3D representation of a scene as well as a neural network trained according to said method and a vehicle comprising such a trained neural network. In particular, the proposed training method comprises: - inputting a first image of the scene acquired by a first camera to the first neural network, said first neural network being configured to: - determine a first 3D representations of the scene based on a representation of the first camera determined by the first neural network from said first image, and a first camera-conditioned depth feature map of the scene determined by the first neural network based on said representation of the first camera; minimizing a loss LλMSE1 which is a difference between said first 3D representation of the scene and a ground truth 3D representation of the scene.