Pose Estimation via Part-Feature Network Message Passing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Pose estimation in images is challenging due to variations in limb orientation, clothing, viewpoints, background clutters, truncation, and occlusion, which lead to loss of detailed information necessary for accurate body part localization.
Innovation Solution
A method that extracts part-feature maps from images using a Convolutional Neural Network (CNN), specifically a Visual Geometry Group (VGG) net with reduced pooling layers, and passes messages between these maps through a part-feature network to update and combine them into score maps, preserving rich inter-part information for improved pose estimation accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If appearance score maps are used to estimate pose locations, then the pose estimation process can be simplified, but detailed inter-part information such as limb orientation, rotation, and occlusion status is lost
Solution Approach 1:
The patent segments the appearance score map into multiple part-feature maps, each corresponding to a specific body part. This segmentation allows preservation of detailed information for each part while maintaining the overall pose estimation framework. Each part_feature map captures local characteristics such as limb orientation, rotation, and occlusion status independently.
Solution Approach 2:
The patent transitions from a single-dimensional appearance score map to a multi-dimensional part_feature map structure. By adding the dimension of part-specific feature extraction, the system preserves inter-part information while maintaining pose estimation capability. The part_feature maps include additional channels for capturing orientation, rotation, and occlusion information.
2Measurement precision
If detailed inter-part information is preserved at the feature level, then pose estimation accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent divides the computational task into separate part_feature map processing streams. Each body part is processed independently through its own feature extraction and message passing mechanism, which reduces the computational burden compared to processing the entire image as a single unit while maintaining high accuracy.
Solution Approach 2:
The patent introduces part_feature maps as intermediary structures between the input image and the final pose estimation. These intermediaries process and refine information locally before being integrated into the overall pose solution, reducing computational complexity by avoiding direct processing of the entire image at high resolution.
3Productivity
If multiple body parts are processed independently, then computational efficiency is improved, but structural learning among body parts becomes less effective
Solution Approach 1:
The patent implements a feedback mechanism through message passing between part_feature maps. Each part's feature map sends messages to its neighboring parts, allowing structural information to flow bidirectionally. This feedback loop enables independent processing of each part while maintaining effective structural learning through iterative message exchange.
Solution Approach 2:
The patent merges the independently processed part_feature maps through message passing and integration operations. After each part is processed independently, the system combines the results by passing messages between adjacent parts, thereby maintaining computational efficiency while restoring structural learning capability through the integration step.
Data Source
AI summary
A method and a system for pose estimation are provided. The method includes: extracting a plurality of sets of part-feature maps from an image, each set of the extracted part-feature maps encoding the messages for a particular body part and forming a node of a part-feature network; passing a message of each set of the extracted part-feature maps through the part-feature network to update the extracted part-feature maps, resulting in each set of the extracted part-feature maps incorporating the message of upstream nodes; estimating, based on the updated part-feature maps, the body part within the image.