Object Alignment from 2D Image via Triangular Mesh Subdivision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object recognition systems face challenges in identifying objects from 2D images when the object is not in a normalized pose, particularly when the pose variation exceeds five degrees, and current solutions are computationally intensive for depth estimation from a single 2D image.
Innovation Solution
The method involves defining a 2D shape and a 3D shape using triangles, mapping corresponding points, subdividing these shapes into new triangles, and assigning z-coordinates from the 3D shape to the 2D shape to reconstruct a 3D image, allowing for object alignment across various poses by creating a 3D reconstructured shape.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional object recognition systems are used to identify objects in 2-D images, then the system can identify objects in normalized poses, but the system fails to accurately identify objects when pose variations exceed five degrees
Solution Approach 1:
The patent transitions from 2-D image processing to 3-D shape representation by defining both a 2-D shape from the image and a corresponding 3-D shape model. This dimensional transformation allows the system to capture depth information and handle pose variations that cannot be represented in 2-D space, thereby improving both measurement precision and adaptability to different poses.
Solution Approach 2:
The patent changes the parameter representation from 2-D coordinates to 3-D coordinates with z-coordinates assigned to vertices. By transforming the geometric parameters from planar to spatial dimensions, the system can accurately represent objects in various poses while maintaining identification precision.
2Measurement precision
If methodologies respecting non-rigid deformation of face are used to align the face, then the face alignment accuracy improves, but the computational intensity increases significantly
Solution Approach 1:
The patent segments the face into triangular meshes, dividing the complex non-rigid deformation problem into smaller, manageable triangular elements. This segmentation allows for efficient computation of depth coordinates while maintaining accuracy in representing facial expressions and pose variations, thereby reducing computational intensity compared to full-face methodologies.
Solution Approach 2:
The patent creates a 3-D shape model that copies the structural information from the 2-D image and assigns z-coordinates to vertices. This copying approach from 2-D to 3-D space provides an efficient alternative to computationally intensive depth estimation methods, maintaining alignment accuracy while reducing computational energy consumption.
Data Source
AI summary
The present disclosure provides methods, machine readable media, and systems for object alignment from a 2-dimensional (2-D) image of the object. One or more embodiments include defining a 2-D shape in the 2-D image and a 3-dimensional (3-D) shape in a 3-D model of the object, mapping a number of corresponding points on the 2-D and 3-D shapes, defining the 2-D and 3-D shapes with a number of triangles, wherein a number of vertices of the number of triangles correspond to the number of points, subdividing the number of triangles defining the 2-D and 3-D shapes into a plurality of subdivided triangles that include a plurality of new vertices, and reconstructuring a 3-D image from the 2-D image by assigning a number of z-coordinates from the plurality of subdivided triangles of the 3-D shape to the plurality of subdivided triangles of the 2-D shape to create a 3-D reconstructured shape.