Hybrid Pose Estimation Model Filtering Low-Reliability Key Points
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing person pose estimation methods struggle with inaccurate key point detection due to non-existent key points and quantization errors, particularly when key points are densely arranged, leading to reduced accuracy and reliability.
Innovation Solution
A hybrid pose estimation model that filters out key points with low reliability by comparing coordinate values with probability values from a visibility map, using a threshold to remove unreliable joints and align key point distributions with an uncorrelated visibility map.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If heatmap-based estimation method is used to remove non-existent key points, then reliability of key point detection is improved, but quantization errors occur and densely arranged key points cannot be distinguished
Solution Approach 1:
The patent segments the key point detection process into two independent branches: one for detecting key point coordinates and another for generating visibility maps. This segmentation allows each branch to specialize - the coordinate detection branch focuses on precision while the visibility map branch focuses on reliability, resolving the contradiction between the two requirements.
Solution Approach 2:
The patent introduces visibility maps as an intermediary element that mediates between the detected key points and the final pose estimation. The visibility maps provide probability information that helps distinguish reliable key points from unreliable ones, especially for densely arranged key points, without introducing quantization errors.
2Productivity
If heatmap-less regression method is used for coordinate estimation, then processing speed is improved, but unexpected results occur due to non-existent key points
Solution Approach 1:
The patent merges the advantages of both heatmap-based and heatmap-less regression methods by combining coordinate regression (for speed) with visibility map generation (for reliability). The hybrid architecture maintains the fast processing of regression while adding the reliability filtering capability of visibility maps.
Solution Approach 2:
The patent changes the output parameters of the pose estimation model to include not only key point coordinates but also visibility probabilities. This parameter expansion allows the system to maintain fast processing while providing reliability information for filtering non-existent key points.
3Quantity of substance
If all detected key points are used for pose estimation, then completeness of pose information is improved, but accuracy decreases due to inclusion of low-reliability key points
Solution Approach 1:
The patent applies local quality filtering by assigning different reliability scores (visibility probabilities) to different key points based on their individual characteristics. Instead of uniformly treating all key points, the system evaluates each key point's reliability locally and filters them differently, maintaining high accuracy while preserving complete pose information.
Data Source
Figure 1~2
Figure 3
Figure 4~5
AI summary
The present disclosure relates to an apparatus and a method for estimating person pose based on image and method capable of removing low-reliability joint information, and the apparatus for estimating person pose includes a memory storing a hybrid pose estimation model; and a processor estimating a person pose within an image based on the hybrid pose estimation model, in which the processor may obtain image data; input the obtained image data into the pre-trained hybrid pose estimation model to extract coordinate values of key points corresponding to person join locations within the image data and probability values of a visibility map, respectively; check whether the probability value of a joint matching the coordinate value of the key point is equal to or less than a threshold value; and if the probability value of the joint matching the coordinate value of the key point is equal to or less than the threshold value, filter out and remove the key point of the corresponding joint.