3D Object Reconstruction With Part-Based Neural Refinement

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing monocular 3D reconstruction methods face challenges in accurately reconstructing 3D objects from 2D images due to issues such as occlusions, rotations, and global translations, particularly when dealing with complex articulated structures like the human body, and lack effective integration of bottom-up and top-down cues.

Innovation Solution

A synergistic approach combining a part-based model with a multi-task neural network that includes bottom-up cues for feature extraction and top-down model fitting, using a parametric deformable model like SMPL, and an iterative refinement method to align model-based and image-based estimates, leveraging kinematic constraints and geometric losses for improved accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a single CNN-based method is used for 3D reconstruction, then the process is simple, but accuracy is insufficient due to inability to handle occlusions and large pose variations

Engineering Contradiction:
Improve3D joint estimation accuracyVSAvoidreconstruction system complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system segments the 3D reconstruction task into multiple specialized components: a part-based model handles articulated structures and occlusions, a multi-task neural network extracts bottom-up cues, and a refinement method performs top-down model fitting. Each component addresses specific aspects of the reconstruction problem, improving overall accuracy while managing complexity through functional decomposition.

Inventive Principle:
Principle #1Segmentation

2Reliability

If bottom-up and top-down cues are not integrated, then the system is simpler, but robustness to occlusions and pose variations is reduced

Engineering Contradiction:
Improverobustness to occlusions and pose variationsVSAvoidintegration of multiple cues
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The system merges bottom-up cues from a part-based model with top-down cues from a parametric model (e.g., SMPL) through an iterative refinement process. The part-based model provides robustness to occlusions by detecting body parts independently, while the parametric model enforces anatomical constraints. Their integration through iterative optimization combines the strengths of both approaches, improving reliability without excessive complexity.

Inventive Principle:
Principle #5Merging (Combining)

3Measurement precision

If model-based and image-based estimates are not aligned, then processing is faster, but reconstruction accuracy is lower

Engineering Contradiction:
Improvealignment between model-based and image-based estimatesVSAvoiditerative refinement time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system implements feedback through an iterative refinement process where model-based estimates are continuously adjusted based on image-based observations. The parametric model parameters are optimized to minimize the difference between projected model landmarks and detected image landmarks, creating a closed-loop system that progressively improves alignment accuracy.

Inventive Principle:
Principle #23Feedback

4Stability of the object's composition

If kinematic constraints are not applied, then the system is simpler, but anatomical plausibility of the reconstruction is reduced

Engineering Contradiction:
Improveanatomical plausibilityVSAvoidkinematic constraint integration
Core Design Contradiction:
Stability of the object's compositionVSDevice complexity

Solution Approach 1:

The parametric model (e.g., SMPL) serves multiple functions simultaneously: it provides a unified representation of human body shape and pose, enforces anatomical constraints through its kinematic structure, and enables efficient optimization. The kinematic tree structure of the model inherently ensures anatomical plausibility by constraining joint rotations to realistic ranges, while the same model structure also facilitates the optimization process.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP3881285B1Three-dimensional object reconstruction
Publication Date: 2025.07.16 SNAP INC
  • EP3881285B1 patent drawingFigure 1
  • EP3881285B1 patent drawingFigure 2
  • EP3881285B1 patent drawingFigure 3

AI summary

This disclosure relates to reconstructing three-dimensional models of objects from two-dimensional images. According to a first aspect, this specification describes a computer implemented method for creating a three-dimensional reconstruction from a two-dimensional image, the method comprising: receiving a two-dimensional image; identifying an object in the image to be reconstructed and identifying a type of said object; spatially anchoring a pre-determined set of object landmarks within the image; extracting a two-dimensional image representation from each object landmark; estimating a respective three-dimensional representation for the respective two-dimensional image representations; and combining the respective three-dimensional representations resulting in a fused three-dimensional representation of the object.