Face Landmark Coordinates via Spatially Preserving Cascaded Decoders

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing face landmark detection methods face challenges such as lower accuracy due to non-differentiable post-processing operations in heat map regression-based approaches and loss of spatial structure in coordinate regression-based methods, leading to quantization errors and decreased localization performance.

Innovation Solution

A multi-stage feature map generation using a staged convolutional network, followed by a cascaded decoder network with self-attention and deformable attention mechanisms, where initial query embeddings are derived from meaningful image features, reducing the number of decoder layers and utilizing a regression loss function for training.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If heat map regression-based approach is used, then detection accuracy is improved, but non-differentiable post-processing operations prevent end-to-end training

Engineering Contradiction:
Improvedetection accuracyVSAvoidtraining framework complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent replaces the non-differentiable post-processing operations (peak detection, coordinate extraction) with a differentiable coordinate regression module. This substitution allows the entire network to be trained end-to-end using gradient-based optimization, eliminating the need for separate post-processing steps while maintaining detection accuracy.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Extent of automation

If coordinate regression-based approach with fully connected prediction layer is used, then end-to-end training is enabled, but spatial structure of image features is destroyed

Engineering Contradiction:
Improveend-to-end training capabilityVSAvoidlocalization performance
Core Design Contradiction:
Extent of automationVSMeasurement precision

Solution Approach 1:

The patent segments the feature processing into two distinct paths: a spatial structure preservation path that maintains the convolutional feature map structure for accurate localization, and a coordinate prediction path that enables end-to-end training. This segmentation allows both goals to be achieved simultaneously without compromising either.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces an intermediary coordinate prediction module that acts as a bridge between the spatial feature maps and the final landmark coordinates. This intermediary preserves the spatial structure information from the convolutional features while enabling differentiable coordinate regression, thus maintaining both localization performance and end-to-end training capability.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Device complexity

If heat map resolution is reduced, then computational complexity is decreased, but quantization errors occur and prevent further improvement

Engineering Contradiction:
Improvecomputational complexityVSAvoidlandmark detection precision
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent replaces the heat map regression mechanism with direct coordinate regression. This substitution eliminates the need for high-resolution heat maps and subsequent quantization, allowing the model to predict landmark coordinates directly from the feature maps without suffering from resolution-related quantization errors.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentEP4181092B1Method and apparatus with face landmark coordinate prediction
Publication Date: 2025.08.20 SAMSUNG ELECTRONICS CO LTD
  • EP4181092B1 patent drawingFigure 1
  • EP4181092B1 patent drawingFigure 2
  • EP4181092B1 patent drawingFigure 3

AI summary

A method and apparatus with landmark coordinate prediction are provided. The method includes generating a multi-stage feature map for landmarks of a face image through a staged convolutional network, generating an initial query matrix by fully connecting a last-stage feature map in the multi-stage feature map using a fully connected network, where a total number of feature elements in the initial query matrix is equal to a total number of predicted landmarks of the face image, generating a memory feature matrix by flattening and connecting the multi-stage feature map, generating the predicted landmark coordinates by inputting the memory feature matrix and the initial query matrix to a decoder network of plural cascaded decoder networks.