3D Motion Estimation From Monocular Images Using Shape Priors

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for estimating three-dimensional motion flow between two-dimensional images, particularly scene flow, are challenging due to factors like depth ambiguity and the aperture problem, often relying on stereoscopic or RGB-D images and enforcing priors on surface smoothness or motion rigidity, which are difficult to capture in real-world scenarios.

Innovation Solution

A convolutional neural network (CNN) is used to estimate three-dimensional flow by inputting two-dimensional images and a two-dimensional representation of an estimated three-dimensional shape of an object, processing them through multiple layers to generate accurate three-dimensional flow without requiring stereographic or RGB-D images, utilizing a trunk network and correlation layers for better initialization and warping in 3D space.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If stereoscopic or RGB-D images are used with priors on surface smoothness or motion rigidity, then three-dimensional flow estimation can be performed, but the method complexity and device requirements increase significantly

Engineering Contradiction:
Improvethree-dimensional flow estimation accuracyVSAvoidimaging system complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts and removes the requirement for complex stereoscopic or RGB-D imaging systems by developing a method that works with standard monocular two-dimensional images. The solution extracts sufficient three-dimensional motion information through computational techniques rather than relying on complex hardware configurations, thereby reducing device complexity while maintaining estimation capability

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent replaces the mechanical/optical system of stereoscopic or RGB-D cameras with a computational approach using convolutional neural networks. Instead of using multiple cameras or specialized sensors to capture three-dimensional information, the method uses a single two-dimensional image processed through learned computational models to estimate three-dimensional flow, substituting hardware complexity with software intelligence

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Reliability

If priors on surface smoothness or motion rigidity are enforced, then the ill-posed problem can be solved, but the accuracy deteriorates in real-world scenarios where these assumptions do not hold

Engineering Contradiction:
Improveproblem solvabilityVSAvoidthree-dimensional flow estimation accuracy
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent introduces dynamic and flexible priors through the convolutional neural network that can adapt to different scenarios. Rather than enforcing fixed assumptions about surface smoothness or motion rigidity, the learned model dynamically adjusts its expectations based on the input images, allowing it to handle both rigid and non-rigid motions, smooth and complex surfaces, thereby maintaining reliability across diverse real-world conditions

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent changes the parameters of the prior assumptions from fixed constraints (surface smoothness, motion rigidity) to learned, adaptive parameters within the neural network. The network learns optimal prior representations from training data, allowing the priors to change and adapt to different object types, motions, and scenarios, thus maintaining problem solvability while improving accuracy in real-world applications

Inventive Principle:
Principle #35Parameter changes

3Device complexity

If two-dimensional images are used instead of stereoscopic or RGB-D images, then device complexity is reduced, but depth information becomes ambiguous

Engineering Contradiction:
Improveimaging system complexityVSAvoiddepth information
Core Design Contradiction:
Device complexityVSLoss of information

Solution Approach 1:

The patent introduces an intermediary computational process using convolutional neural networks that bridges the gap between two-dimensional images and three-dimensional flow estimation. The network acts as an intermediary that processes the limited two-dimensional information and reconstructs the missing depth and motion information through learned patterns, effectively mediating between the simplified input and the complex output required

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent performs preliminary action by pre-training the convolutional neural network on large datasets of images and corresponding three-dimensional flow annotations. This preliminary training equips the network with prior knowledge about depth cues, motion patterns, and three-dimensional geometry, enabling it to accurately estimate three-dimensional flow from two-dimensional images without requiring complex imaging hardware

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4128154B1Three-dimensional motion estimation
Publication Date: 2026.02.25 HUAWEI TECH CO LTD
  • EP4128154B1 patent drawingFigure 1
  • EP4128154B1 patent drawingFigure 2
  • EP4128154B1 patent drawingFigure 3

AI summary

A method for estimating a three-dimensional flow of an object between a first two-dimensional image comprising an image of the object in a first configuration and a second two-dimensional image comprising an image of the object in a second configuration, the method comprising: inputting the first two-dimensional image, the second two-dimensional image and a two-dimensional representation of an estimated three-dimensional shape of the object into a convolutional neural network; and generating, using the convolutional neural network, the three-dimensional flow of the object between the first two-dimensional image and the second two-dimensional image from the first two-dimensional image, the second two-dimensional image and the two-dimensional representation of the estimated three-dimensional shape of the object.