View-Synthesis Depth Learning With GAN-Based Image Authenticity

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing depth estimation methods for autonomous driving face challenges in generating high-quality synthetic images due to limitations in learning and increased inference time and memory usage, particularly with self-supervised depth estimation methods using CRF or RNN-based approaches.

Innovation Solution

A method and apparatus that utilize a generative adversarial network (GAN) to train a synthetic image model, incorporating a depth network and pose network, to generate high-quality synthetic images by freezing synthetic parameters and applying multiple loss functions for consistency and authenticity, thereby improving depth estimation accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If self-supervised depth estimation methods using CRF or RNN-based approaches are utilized, then depth estimation performance may be improved, but inference time and memory usage drastically increase

Engineering Contradiction:
Improvedepth estimation accuracyVSAvoidinference time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent extracts and removes the CRF and RNN components from the depth estimation pipeline, replacing them with a simplified neural network architecture that achieves comparable accuracy without the computational burden of these complex post-processing steps

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent replaces the mechanical/algorithmic CRF and RNN processing mechanisms with a learned neural network model that directly outputs depth estimates, substituting complex iterative algorithms with a single-pass differentiable computation

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Measurement precision

If self-supervised depth estimation methods using CRF or RNN-based approaches are utilized, then depth estimation performance may be improved, but memory usage drastically increases

Engineering Contradiction:
Improvedepth estimation accuracyVSAvoidmemory usage
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent removes the memory-intensive CRF and RNN components from the system, replacing them with a lightweight neural network architecture that maintains accuracy while significantly reducing memory consumption

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent employs a simplified neural network model that can be efficiently trained and discarded/retrained as needed, replacing the need for maintaining complex CRF/RNN structures in memory during inference

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

3Productivity

If a loss function based on human experience and knowledge is utilized, then the model may converge easier, but learning high-quality image synthesis is limited

Engineering Contradiction:
Improvemodel convergence speedVSAvoidimage synthesis quality
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent changes the loss function parameters from hand-crafted metrics based on human experience to learned metrics that automatically adapt to the specific dataset and task, enabling both fast convergence and high-quality synthesis through data-driven optimization

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent implements a self-supervised learning framework where the model learns to synthesize images and evaluate its own output quality without external human annotations, allowing the system to improve synthesis quality autonomously while maintaining efficient convergence

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP4651082A1Method and device for learning depth estimation based on view synthesis
Publication Date: 2025.11.19 HYUNDAI MOTOR CO LTD
  • EP4651082A1 patent drawingFigure 1~2
  • EP4651082A1 patent drawingFigure 3
  • EP4651082A1 patent drawingFigure 4(a)~4(b)

AI summary

A method for controlling autonomous driving of a vehicle is introduced. The method may comprise, training, based on an inference depth and an inference pose, a synthetic image model for generating a synthetic image, generating, based on the synthetic image, a first virtual image to be associated with the original image, generating, based on the original image, a second virtual image, training a generative adversarial network (GAN) for determining, based on the original image, authenticity of the first virtual image and the second virtual image, training, based on the trained GAN, a depth network, wherein the trained GAN outputs a determination of the authenticity of the first virtual image, outputting, based on the trained depth network, signal, and controlling, based on the signal, autonomous driving of the vehicle.