3D Human Model Generation With Part-Attentive Pose Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods struggle to generate a three-dimensional human model accurately from a single image, particularly when body parts are occluded, and lack effective utilization of kinematic joint relationships for pose prediction.
Innovation Solution
A method and device that generate a human model by creating feature maps for body centers, parts, and parameters, using a part-attentive feature generation and adjustment process, including matrix operations to derive pose and shape parameters, considering kinematic joint relationships.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If a single image is used to generate a three-dimensional human model, then the input data requirement is reduced, but the accuracy of pose parameter derivation deteriorates due to occlusions and limited information
Solution Approach 1:
The patent divides the human body into multiple body parts (head, torso, arms, legs) and generates separate feature maps for each part. This segmentation allows the system to process and analyze each body part independently, improving pose parameter accuracy even when parts are occluded in the single input image
Solution Approach 2:
The patent transforms the two-dimensional single image input into three-dimensional pose parameters by generating multi-channel feature maps that encode spatial, semantic, and contextual information. This dimensional transformation enables accurate 3D human model generation from limited 2D input data
2Measurement precision
If feature maps are generated for each body part, then pose prediction accuracy improves, but the device complexity increases due to multiple processing units
Solution Approach 1:
The patent employs a single neural network model that performs multiple functions: generating body center maps, part index maps, body part maps, and parameter maps simultaneously. This multi-functional approach improves pose prediction accuracy without proportionally increasing device complexity, as one unified structure handles all feature map generation tasks
Solution Approach 2:
The patent combines multiple feature maps (body center map, part index map, body part map, parameter map) into a unified part-attentive feature representation. This merging process integrates information from different sources to improve pose prediction accuracy while avoiding the need for separate complex processing units for each map type
3Reliability
If kinematic joint relationships are considered in pose parameter derivation, then the reliability of human model generation improves, but the computational complexity increases
Solution Approach 1:
The patent pre-establishes the kinematic joint relationship structure and body part hierarchy before processing input images. By preparing the kinematic model and joint constraints in advance, the system can reliably enforce physical constraints during pose parameter derivation without adding significant computational complexity during actual image processing
Data Source
AI summary
A method of generating a human model according to the present disclosure may include generating a plurality of feature maps from an input image, wherein the plurality of feature maps include a body center map, a part index map, a body part map and a parameter map, generating a part-attentive feature configured with feature maps for each body part based on the part index map and the body part map, readjusting the part-attentive feature based on the parameter map and generating a pose parameter based on the readjusted part-attentive feature.