Concatenated Pyramid Network for Multi-Person Pose Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current multi-person pose estimation methods in computer vision are computationally expensive and inefficient, especially when dealing with multiple human figures in an image, as they often require multiple stages of prediction and image detection at various scales.
Innovation Solution
A computing system utilizing a concatenated pyramid network as a backbone and parallel head neural networks for single-stage, single-scale prediction, which processes input images to output keypoint heatmaps and part affinity fields, allowing for efficient linking of keypoints into virtual skeletons.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple stages of prediction and image detection at various scales are used for multi-person pose estimation, then measurement precision is improved, but productivity deteriorates due to increased processing time and computational cost
Solution Approach 1:
The network is segmented into a backbone for feature extraction and multiple parallel heads for different prediction tasks (keypoint detection, part affinity fields, objectness scores). This segmentation allows simultaneous execution of multiple functions in a single stage, improving both precision and speed
Solution Approach 2:
The method transforms the problem from spatial multi-scale detection to a single-scale detection in feature map space by processing images at one scale and using learned features to capture multi-scale information. This dimensional transformation eliminates the need for multiple processing stages while maintaining accuracy
2Manufacturing precision
If multiple stages of prediction are used for multi-person pose estimation, then manufacturing precision is improved, but device complexity increases
Solution Approach 1:
Multiple prediction functions (keypoint detection, part affinity field estimation, objectness scoring) are merged into a single unified network pass with parallel heads. This combining approach maintains high precision while reducing architectural complexity compared to cascaded multi-stage systems
Solution Approach 2:
The backbone network serves multiple functions simultaneously: feature extraction for keypoint detection, part affinity field generation, and objectness scoring. This multi-functionality reduces overall system complexity while maintaining high precision across all tasks
3Measurement precision
If image processing is performed in multiple stages, then measurement precision is improved, but loss of time increases
Solution Approach 1:
The backbone network performs preliminary feature extraction that captures multi-scale information in advance, eliminating the need for subsequent multi-scale processing stages. This preliminary action maintains detection precision while reducing overall processing time
Solution Approach 2:
The network performs all prediction tasks (keypoint detection, part affinity fields, objectness scoring) in a continuous single-stage process rather than discrete multi-stage processing. This continuity maintains accuracy while minimizing processing time delays
Data Source
AI summary
A computing system is provided. The computing system includes a processor configured to execute a convolutional neural network that has been trained, the convolutional neural network including a backbone network that is a concatenated pyramid network, a plurality of first head neural networks, and a plurality of second head neural networks. At the backbone network, the processor is configured to receive an input image as input and output feature maps extracted from the input image. The processor is configured to: process the feature maps using each of the first head neural networks to output corresponding keypoint heatmaps; process the feature maps using each of the second head neural networks to output corresponding part affinity field heatmaps; link the keypoints into one or more instances of virtual skeletons using the part affinity fields; and output the instances of the virtual skeletons.