Facial Landmark Detection with Spatio‐temporal Modeling
Romain Belmonte, Pierre Tirilly, Ioan Marius Bilasco, Nacim Ihaddadene, Chaabane Djéraba · 2022
This chapter describes current possibilities for spatio-temporal modeling in a broader context. It reviews both hand-crafted features and deep learning approaches. The chapter introduces solution proposal. It presents experimental protocol, implementation details and results with their analysis. The chapter provides experiments on two datasets, 300VW and SNaP-2DFe, in order to evaluate the results obtained and to compare them with state-of-the-art approaches. Spatio-temporal modeling is generally built upon image-based solutions that are extended to video. The chapter describes the architectures developed to extend the connectivity of convolutional neural networks-based landmark detectors to include local motion through early connectivity. It analyzes the performance of each model in terms of speed, size and number of parameters. The chapter explores the complementarity between local and global motion has also been subject to new experiments.