3D Human Reconstruction with Occlusion Handling

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for reconstructing humans in 3D face challenges such as missing color and depth information due to occlusions, leading to loss of detail and incomplete representations.

Innovation Solution

A method that involves capturing multiple color and depth images from different viewpoints, using image segmentation to identify occluded regions, generating a point cloud, discounting points corresponding to occluding items, and adjusting a parametric model to reconstruct the human subject with occluded parts included.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If color and depth images are captured to form a point cloud, then detail preservation is improved, but completeness deteriorates due to occlusions

Engineering Contradiction:
Improvedetail preservationVSAvoidcompleteness
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The method segments the point cloud into multiple components: visible surface points, occluded region points (identified through depth discontinuities and boundary analysis), and inferred points. This segmentation allows different processing strategies for different regions, preserving detail where visible while inferring complete geometry where occluded.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The method transitions from 2D image space to 3D point cloud space, then to a complete 3D mesh representation. By working in three dimensions and analyzing depth maps, the system can identify occluded regions and infer their geometry, recovering information that is lost when staying in 2D image space.

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

2Productivity

If deformable model fitting is used, then reconstruction speed is improved, but detail completeness deteriorates due to occlusions by hair or clothing

Engineering Contradiction:
Improvereconstruction speedVSAvoiddetail completeness
Core Design Contradiction:
ProductivityVSLoss of information

Solution Approach 1:

The method performs preliminary analysis of the point cloud to identify occluded regions before final mesh generation. By detecting depth discontinuities and boundary artifacts in advance, the system can prepare appropriate filling strategies, ensuring that occluded regions like hair and clothing are properly reconstructed before the final model is created.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The method introduces an intermediary processing stage between point cloud generation and final mesh creation. This intermediate stage includes occlusion detection, region classification, and selective point filtering, which prepares the data for complete reconstruction while maintaining the efficiency of automated processing.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3850588B1Method and system for generating a 3D reconstruction of a human
Publication Date: 2025.02.12 SONY INTERACTIVE ENTERTAINMENT LLC
  • EP3850588B1 patent drawingFigure 1
  • EP3850588B1 patent drawingFigure 2
  • EP3850588B1 patent drawingFigure 3~4

AI summary

A method of generating a three-dimensional (3D) reconstruction of a human is provided. The method comprises obtaining at least one colour image and corresponding depth image of a scene comprising a human subject that is at least partially occluded by one or more items. The regions in the at least one colour image corresponding to a human subject are identified. A point cloud of the scene is generated based on the at least one colour image and corresponding depth image. One or more parameters of a parametric model are then adjusted based on the regions of the point cloud identified as corresponding to the human subject. The adjusted parametric model is then used to generate a 3D reconstruction of the human subject; wherein the parametric model provides an estimate for points in the point cloud corresponding to the occluded parts of the human subject. A corresponding system is also provided.