3D Occlusion Reconstruction With CoDeep GANs for Point Clouds

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Occlusions in computer-simulated environments, particularly in point cloud data, pose challenges due to training instability, lack of data diversity, and quality control in Generative Adversarial Networks (GANs), leading to issues like mode collapse and biased generation.

Innovation Solution

A method and system using Neural Networks, specifically Co-evolutionary Deep Generative Adversarial Networks (CoDeep GANs), to detect occluded regions, optimize model parameters, and iteratively generate and refine GAN models to reconstruct occluded areas in 3D multimedia data, employing techniques like crossover, speciation, and mutation to enhance model performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If Generative Adversarial Networks (GANs) are used to address occlusion in point clouds, then occluded regions can be reconstructed, but training instability and mode collapse occur

Engineering Contradiction:
Improveocclusion reconstruction reliabilityVSAvoidtraining stability
Core Design Contradiction:
ReliabilityVSStability of the object's composition

Solution Approach 1:

The patent segments the point cloud data into visible and occluded regions, allowing the GAN to focus specifically on reconstructing occluded areas rather than processing entire scenes. This segmentation approach stabilizes training by reducing the complexity of the reconstruction task and preventing mode collapse through targeted generation.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies preliminary processing steps including point cloud registration, segmentation, and feature extraction before feeding data to the GAN. These preliminary actions prepare the data in an optimized format that stabilizes subsequent GAN training and improves reconstruction reliability.

Inventive Principle:
Principle #10Preliminary action

2Adaptability or versatility

If GANs are used for point cloud augmentation, then diverse data can be generated, but complexity of the model increases

Engineering Contradiction:
Improvedata diversityVSAvoidmodel complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent implements local quality by applying different processing strategies to different regions of the point cloud. Visible regions undergo standard processing while occluded regions receive specialized GAN-based reconstruction. This localized approach maintains data diversity without requiring the entire model to be highly complex.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent introduces intermediary components including feature extraction networks and loss function mechanisms that mediate between the input point cloud and the GAN generator. These intermediaries simplify the overall model complexity by preprocessing data and guiding the GAN's generation process.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Loss of time

If GANs are trained on limited point cloud data, then training can be completed faster, but quality of generated point clouds deteriorates

Engineering Contradiction:
Improvetraining timeVSAvoidgenerated point cloud quality
Core Design Contradiction:
Loss of timeVSManufacturing precision

Solution Approach 1:

The patent implements periodic action through iterative training cycles with alternating optimization of the generator and discriminator networks. This periodic update mechanism allows the model to progressively improve quality even with limited data, balancing training speed and generation quality through structured training phases.

Inventive Principle:
Principle #19Periodic action

Solution Approach 2:

The patent incorporates multiple feedback mechanisms including adversarial loss from the discriminator, reconstruction loss comparing generated points with ground truth occluded regions, and regularization terms. These feedback signals guide the GAN to maintain high quality generation even when trained on limited datasets.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP4704039A1A method and system for managing occlusions in scenes rendered in a computer-simulated environment
Publication Date: 2026.03.04 SIEMENS AG
  • EP4704039A1 patent drawingFigure 1
  • EP4704039A1 patent drawingFigure 2
  • EP4704039A1 patent drawingFigure 3

AI summary

The present invention relates a method and system for managing occlusions in one or more scenes rendered in a computer-simulated environment (120). The method comprises processing 3D multimedia data of scenes from data acquisition devices (104) to detect occluded regions using Neural Networks (NNs). The method comprises determining model optimization parameters based on these occlusions and generating primary Generative Adversarial Network (GAN) models tailored to these regions. The fitness of these GAN models is evaluated by simulating occluded data. The method comprises iteratively generating secondary GAN models until optimal criteria are met. The best-performing GAN models are identified and used to reconstruct the occluded regions within the 3D multimedia data. This reconstructed data is superimposed onto the original scenes to create an augmented 3D visualization. The enhanced 3D scenes are displayed on a user interface associated with an electronic device (110).