GAN-Generated Raw Depth Maps for Realistic Depth Completion Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing depth sensors in industrial settings produce raw depth maps with large errors and loss of information due to optical reflections, deflections, and refractions, especially with reflective metallic objects, necessitating cumbersome and costly real-world data acquisition for training depth completion models.

Innovation Solution

Generate synthetic raw depth maps from CAD data using a generative adversarial network (GAN) to create training data for depth completion models, incorporating a marginal preservation loss to maintain geometric consistency and avoid object removal, and use these synthetic data alongside real raw depth maps for training.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If real-world depth sensor measurements are used to acquire training data, then the training data reflects real-world conditions, but the data acquisition process becomes cumbersome and costly

Engineering Contradiction:
Improverealism of training dataVSAvoiddata acquisition time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent uses a generator model to create synthetic copies of real-world depth maps from synthetic dense depth maps. These synthetic raw depth maps replicate the characteristics of real sensor measurements without requiring actual physical data collection, thus resolving the contradiction between data realism and acquisition efficiency

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The generator model transforms synthetic dense depth maps into synthetic raw depth maps by applying parameter changes that simulate sensor measurement characteristics including noise patterns, reflection artifacts, and incomplete sampling. This transformation enables realistic training data generation without physical data collection

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If depth sensors are used to acquire raw depth maps in industrial settings, then direct 3D geometry measurement is obtained, but large errors and information loss occur due to optical reflections and refractions

Engineering Contradiction:
Improvedepth measurement accuracyVSAvoiddepth information completeness
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The patent applies preliminary action by using a depth completion model to process and correct raw depth maps before they are used for training. This preliminary processing step compensates for sensor errors and recovers lost information, improving both measurement precision and information completeness

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The depth completion model serves as an intermediary between the flawed sensor measurements and the training data requirements. It mediates by transforming erroneous raw depth maps into corrected depth information, eliminating the direct impact of optical reflections and refractions

Inventive Principle:
Principle #24Intermediary (Mediator)

3Manufacturing precision

If synthetic dense depth maps are generated from CAD data, then ideal geometric information is obtained, but the data lacks the characteristics of real sensor measurements

Engineering Contradiction:
Improvegeometric accuracyVSAvoidrealism of depth maps
Core Design Contradiction:
Manufacturing precisionVSReliability

Solution Approach 1:

The generator model creates copies of synthetic dense depth maps that incorporate real sensor measurement characteristics. By copying and transforming the ideal geometric data through a learned mapping, the system achieves both geometric accuracy and measurement realism simultaneously

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4141780B1Method and device for generating training data to generate synthetic real-world-like raw depth maps for the training of domain-specific models for logistics and manufacturing tasks
Publication Date: 2025.07.09 ROBERT BOSCH GMBH
  • EP4141780B1 patent drawingFigure 1
  • EP4141780B1 patent drawingFigure 2
  • EP4141780B1 patent drawingFigure 3

AI summary

The present invention relates to a computer-implemented method for providing training data for training of a data-driven depth completion model as a machine-learning model, wherein the depth completion model is to be trained to generate dense depth maps from sensor acquired raw depth maps, comprising the steps of: - providing (S1) multiple synthetic dense depth map data items (Dsyn) from CAD data of various synthetic scenes; - providing (S2) multiple real raw depth map data items (Draw_real) obtained from real-world depth sensor measurements of real-world scenes; - training (S3) a generative model (10) for obtaining a trained generator model (11) for generating generated raw depth map data items (Draw_syn) from the synthetic dense depth map data items (Dsyn); - applying (S4) the trained generator model (11) to generate training data from provided synthetic dense depth map data (Dsyn).