3D Object Reconstruction With Higher-Order Function Networks

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing 3D object representation methods, such as voxel-based and partition-based approaches, are inefficient in terms of memory and computation, and point set representations struggle with adapting to varied-resolution predictions.

Innovation Solution

Representing 3D objects using higher-order function networks, specifically through a neural network with mapping functions that map points from a canonical space to define the 3D object, allowing for accurate reconstruction and interpolation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If voxel-based approaches are used to represent 3D objects, then the ability to extend CNN based approaches to 3D world is improved, but memory efficiency and computation time deteriorate due to inherent sparsity of 3D surfaces

Engineering Contradiction:

Improveability to extend CNN based approaches to 3D worldVSAvoidmemory efficiency

Core Design Contradiction:

Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The patent extracts only the essential surface information of 3D objects using neural implicit representations, eliminating the need to represent entire volumetric spaces with voxels. This extracts the core geometric features while discarding redundant empty space representations, achieving both CNN compatibility and memory efficiency.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes the representation parameters from discrete voxel grids to continuous implicit functions defined by neural networks. This parameter transformation allows the system to represent 3D surfaces with far fewer parameters while maintaining the ability to perform convolutional operations through differentiable sampling.

Inventive Principle:

Principle #35Parameter changes

2Adaptability or versatility

If voxel-based approaches are used to represent 3D objects, then the ability to extend CNN based approaches to 3D world is improved, but computation time deteriorates due to inherent sparsity of 3D surfaces

Engineering Contradiction:

Improveability to extend CNN based approaches to 3D worldVSAvoidcomputation time

Core Design Contradiction:

Adaptability or versatilityVSLoss of time

Solution Approach 1:

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent performs computation only on sampled points along ray paths rather than processing entire voxel grids. This partial action approach computes only where necessary (at surface intersections) rather than exhaustively processing all volumetric elements, significantly reducing computation time.

Inventive Principle:

Principle #16Partial or excessive action

3Quantity of substance

If partition-based approaches such as octrees are used, then space efficiency is improved, but implementation complexity and query efficiency deteriorate

Engineering Contradiction:

Improvespace efficiencyVSAvoidimplementation complexity

Core Design Contradiction:

Quantity of substanceVSDevice complexity

Solution Approach 1:

The patent replaces the mechanical hierarchical partitioning structure of octrees with a smooth neural field representation. This substitution eliminates complex data structures and partitioning logic while maintaining space efficiency through the compact neural network parameters that define the entire 3D surface implicitly.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

4Quantity of substance

If point set representations are used, then storage and computational burden is reduced, but adaptability to arbitrary- or varied-resolution predictions deteriorates

Engineering Contradiction:

Improvestorage burdenVSAvoidadaptability to arbitrary- or varied-resolution predictions

Core Design Contradiction:

Quantity of substanceVSAdaptability or versatility

Solution Approach 1:

The patent creates a dynamic representation where the neural implicit function can be sampled at any resolution level on demand. The same compact neural network parameters can generate point clouds at low resolution or detailed surface meshes at high resolution, making the representation adaptable to varying prediction requirements without storing multiple resolutions.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The patent creates a universal representation that serves multiple functions: it can generate point clouds for storage efficiency, produce high-resolution surfaces for detailed predictions, and enable gradient-based optimization for arbitrary resolution outputs. A single neural implicit function replaces multiple specialized data structures.

Inventive Principle:

Principle #6Universality (Multi-functionality)

Data Source

PatentEP3953894B1Higher-order function networks for learning composable three-dimensional (3D) object and operating method thereof

Publication Date: 2025.12.10 SAMSUNG ELECTRONICS CO LTD

Figure 1
Figure 2A
Figure 2B

AI summary

An apparatus for representing a three-dimensional (3D) object, the apparatus includes a memory storing instructions, and a processor configured to execute the instructions to transmit a two-dimensional (2D) image to an external device, based on the 2D image being transmitted, receive, from the external device, mapping function parameters that are obtained using a first neural network, set a mapping function of a second neural network, based on the received mapping function parameters, and based on 3D samples, obtain the 3D object corresponding to the 2D image, using the second neural network of which the mapping function is set.