Key Point Positioning via Dual Feature Map Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current human key point positioning methods using deep neural networks and heat maps often result in errors due to the limitations of directly relying on heat maps for accurate positioning of key points in images.

Innovation Solution

A method that extracts two independent feature maps from a deep neural network, one for rough positioning using large-scale global semantic features and another for accurate positioning using small-scale local semantic features, allowing for improved accuracy by combining these features to determine the final position of key points.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If a deep neural network with heat map is used for key point positioning, then the implementation is simple and direct, but the positioning accuracy is insufficient

Engineering Contradiction:
Improveimplementation simplicityVSAvoidkey point positioning accuracy
Core Design Contradiction:
Ease of manufactureVSMeasurement precision

Solution Approach 1:

The patent segments the key point positioning task into two distinct stages: rough positioning using a first feature map with large-scale global semantic features, and accurate positioning using a second feature map with small-scale local semantic features. This segmentation allows each stage to focus on specific aspects of the positioning problem, improving overall accuracy while maintaining implementation feasibility through modular processing

Inventive Principle:
Principle #1Segmentation

2Ease of operation

If only large-scale global semantic features are used for positioning, then the processing is straightforward, but the positioning accuracy is limited

Engineering Contradiction:
Improveprocessing simplicityVSAvoidpositioning accuracy
Core Design Contradiction:
Ease of operationVSMeasurement precision

Solution Approach 1:

The patent introduces a new dimension in feature processing by extracting both large-scale global semantic features and small-scale local semantic features from the deep neural network. This multi-dimensional feature extraction approach enriches the information available for positioning, enabling more accurate key point detection while maintaining operational simplicity through systematic feature integration

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Measurement precision

If only small-scale local semantic features are used for positioning, then local detail accuracy is good, but global context is lost

Engineering Contradiction:
Improvelocal positioning accuracyVSAvoidglobal semantic context
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The patent merges the results from two separate feature maps: the first feature map providing global semantic context through large-scale features, and the second feature map providing local detail accuracy through small-scale features. By combining these complementary information sources, the system achieves both global context retention and local positioning precision without losing important semantic information

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP3869402B1Method and apparatus for positioning a key point, device, storage medium and computer program product
Publication Date: 2023.05.17 BEIJING BAIDU NETCOM SCI & TECH CO LTD
  • EP3869402B1 patent drawingFigure 1~2
  • EP3869402B1 patent drawingFigure 3
  • EP3869402B1 patent drawingFigure 4~5

AI summary

Embodiments of the present disclosure disclose a method and apparatus for positioning a key point, a device, a storage medium and a computer program product, and relate to the technical fields of artificial intelligence, image processing and deep learning. A specific embodiment of the method includes: extracting a first feature map and a second feature map of a to-be-positioned image, the first feature map and the second feature map being different feature maps; determining, based on the first feature map, an initial position of a key point in the to-be-positioned image; determining, based on the second feature map, an offset of the key point; and adding the initial position of the key point with the offset of the key point to obtain a final position of the key point. Rough positioning and accurate positioning are performed on the key point by using different feature maps respectively, thereby improving the abilities of the feature maps to learn different types of features concerned by the rough positioning and the accurate positioning, and ultimately improving the accuracy of key point positioning.