Abstract:In this paper, we introduce a novel method for detecting facial landmarks, termed the spatial fusion regression augmentation network (SFRA), which aims to improve the spatial generalization capabilities of coordinate based regression techniques. The SFRA consists of three components, namely, the landmark feature enhancement network (LFE), spatial relationship modeling network (SRM), and landmark coordinate regression (LCR) network. The LFE is capable of incorporating prior information into the extracted landmark features, thus boosting their representational capacity. The SRM is capable of understanding the local interaction and global interaction among the landmark patches before they are passed to the LCR to predict the coordinates of the landmarks. Then, the output from each predicted coordinates from the LCR feeds into the subsequent stage's LFE to improve feature representation, creating a pyramid-like multistage structure that refines details progressively from a coarse to a fine level, ultimately achieving precise facial landmark predictions. The proposed method was rigorously tested on the 300W, WFLW and COFW datasets, achieving normalized mean error (NME) metrics of 3.07%, 4.11% and 3.29% respectively, thus placing it to be the best performing coordinate based regression methods; its performance when compared with the best heatmap based regression methods, is mixed, in that on the WFLW dataset, it is better, on the 300W dataset it is slightly worse, and on the COFW dataset, it is better if the normalization factor in the NME metric is based on inter-ocular distances, while using inter-pupil distances, it is slightly worse than the best method in the heatmap based regression methods. Lastly, ablation studies are performed to confirm the efficacy of the proposed approach.

Matrix Factorization with Landmarks for Spatial Data.

Feature Extraction Via Multi-View Non-Negative Matrix Factorization with Local Graph Regularization.

Privacy Preserving Point-of-Interest Recommendation Using Decentralized Matrix Factorization.

GeoMF++

Self-Paced Learning for Matrix Factorization

Sparse Concept Discriminant Matrix Factorization for Image Representation.

Accelerating Locality Preserving Nonnegative Matrix Factorization

Matrix Factorization for Spatio-Temporal Neural Networks with Applications to Urban Flow Prediction

Spatial Factor Modeling: A Bayesian Matrix-Normal Approach for Misaligned Data

Graph regularized non-negative matrix factorization with sparse coding

A 3D Statistical Facial Feature Model and Its Application on Locating Facial Landmarks

Toward Efficient Image Representation: Sparse Concept Discriminant Matrix Factorization

Spatially correlated nonnegative matrix factorization

Locality-constrained Concept Factorization

A modified flexible spatiotemporal data fusion model

A locally weighted sparse graph regularized Non-Negative Matrix Factorization method.

SFRA: spatial fusion regression augmentation network for facial landmark detection

Spatially correlated nonnegative matrix factorization for image analysis

Locality Preserving Nonnegative Matrix Factorization.

Matrix Factorization with Column L0-Norm Constraint for Robust Multi-subspace Analysis

Regularized Nonnegative Matrix Factorization with Adaptive Local Structure Learning.