CNN Pose Estimation Using Directional Feature Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for estimating two-dimensional vector fields using neural networks face challenges in accurately training the networks, leading to suboptimal accuracy in detecting the relative positional relationships between parts of a subject.
Innovation Solution
An image processing apparatus and method utilizing a convolutional neural network (CNN) that generates feature maps indicating the approximate direction of relative position vectors, trained using loss functions specific to both the relative position vectors and feature maps, to enhance the detection of subject parts and their positional relationships.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a neural network is used to estimate two-dimensional vector fields directly, then the system can detect relative positional relationships between body parts, but the training difficulty increases and estimation accuracy deteriorates
Solution Approach 1:
The patent segments the vector field estimation task into two separate detection tasks: (1) detecting relative position vectors between body parts, and (2) detecting feature maps indicating approximate directions. This segmentation allows each task to be optimized independently with appropriate loss functions, improving overall accuracy while managing training complexity through modular approach
Solution Approach 2:
The patent introduces feature maps as an intermediary representation between the input image and the final relative position vector estimation. These feature maps serve as intermediate outputs that capture directional information, facilitating the training process by providing intermediate targets and gradient flow paths that improve convergence and final accuracy
2Reliability
If the neural network focuses on estimating relative position vectors directly, then the system can obtain positional information, but the association reliability between body parts deteriorates when estimation is inaccurate
Solution Approach 1:
The patent implements feedback mechanisms through dual loss functions that provide continuous guidance during training. The first loss function supervises the relative position vector predictions, while the second loss function supervises the feature map predictions. This multi-feedback approach ensures the network learns reliable associations by receiving corrective signals from multiple sources, maintaining robustness even when individual estimates have errors
Data Source
AI summary
An image processing apparatus that estimates a vector representing a relative positional relationship between parts of a subject using machine learning is disclosed. The image processing apparatus, using a convolutional neural network (CNN), detects, from an image, a first part and a second part of a specific subject, and a relative position vector from the first part to the second part. The CNN is configured to generate a feature map indicating an approximate direction of the relative position vector. The image processing apparatus trains the CNN using a loss function pertaining to the relative position vector and a loss function pertaining to the feature map.