Circuitous Pyramid Network for Multi-Person Pose Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The multi-person pose recognition technology faces a bottleneck in feature propagation, leading to insufficient precision in bottom-up recognition methods, which affects the accuracy of pose recognition.
Innovation Solution
A circuitous pyramid network is employed, utilizing phases connected in parallel with downsampling and upsampling networks and residual connection layers to facilitate feature propagation and fusion, avoiding convolution operations and enhancing feature extraction across different scales.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If bottom-up recognition method is used, then processing efficiency is improved, but precision is insufficient
Solution Approach 1:
The network is divided into multiple phases (first phase, second phase, etc.), each phase containing downsampling and upsampling networks. This segmentation allows progressive feature extraction at different scales while maintaining processing efficiency, resolving the contradiction between efficiency and precision.
Solution Approach 2:
The patent introduces a multi-phase hierarchical structure that adds a temporal/dimensional dimension to feature processing. By processing features through multiple phases with increasing abstraction levels, the system achieves both efficient processing and high precision pose recognition.
2Measurement precision
If stacked hourglass network is used, then precision is improved, but feature propagation bottleneck occurs
Solution Approach 1:
The feature propagation path is segmented into multiple phases with distinct downsampling and upsampling networks. Each phase handles specific scale features independently, reducing the bottleneck effect in traditional stacked hourglass networks while maintaining precision.
Solution Approach 2:
The downsampling networks perform preliminary feature extraction and abstraction before the upsampling networks reconstruct the pose information. This preliminary action at multiple phases prepares features in advance, facilitating smoother propagation and reducing computational bottlenecks.
3Loss of information
If convolution operation is used for feature propagation, then feature extraction is performed, but propagation bottleneck is formed
Solution Approach 1:
The feature propagation is segmented into parallel downsampling and upsampling paths across multiple phases. This segmentation allows simultaneous feature extraction and propagation without the sequential bottleneck of traditional convolution operations, improving both information retention and propagation speed.
Solution Approach 2:
The network dynamically processes features through multiple phases where each phase adapts the level of abstraction. This dynamic multi-phase processing enables efficient feature propagation while maintaining extraction capability, overcoming the static bottleneck of conventional convolution-based approaches.
Data Source
Figure 1~2
Figure 3
Figure 4~5
AI summary
The present application discloses a method and an apparatus for recognizing the postures of multiple persons, an electronic device, and a storage medium. The method for recognizing the postures of multiple persons includes: acquiring an image to be recognized; constructing a circuitous hierarchical network, the circuitous hierarchical network comprising several stages in parallel, each stage comprising various layers of a down-sampling network, various layers of an up-sampling network, and a first residual connection layer connected between the various layers of the up-sampling network and the down-sampling network, different stages being connected by means of a second residual connection layer; traversing the stages of the circuitous hierarchical network, comprising performing the following processing: in the feature map extraction performed in the current stage, performing, by means of the first residual connection layer, feature propagation between the layers of the down-sampling network and the layers of the up-sampling network in the current stage, so as to obtain an output feature map of the current stage; performing, by means of the second residual connection layer, feature propagation between the layers of the up-sampling network in the current stage and the layers of the down-sampling network in a later stage, so as to perform feature map extraction corresponding to the later stage, until the traversal of the stages in the circuitous hierarchical network is completed, taking the output feature map of the last stage as the feature map corresponding to said image; and recognizing the postures of multiple persons according to the feature map corresponding to said image, so as to obtain a posture recognition result of said image.