PROJECT TITLE :

Attended End-to-end Architecture for Age Estimation from Facial Expression Videos

ABSTRACT:

The most difficult aspects of estimating age from facial expression films are modeling the static face appearance as well as capturing the temporal facial dynamics. Traditional approaches to this topic concentrate on handcrafting characteristics in order to investigate the discriminative information contained in facial appearance and dynamics separately. This necessitates a high level of feature refinement and framework design. In this research, we introduce Spatially-Indexed Attention Model (SIAM), an end-to-end architecture for age estimation that can simultaneously learn both the appearance and dynamics of age from raw videos of facial expressions. To simulate the temporal dynamics, we use convolutional neural networks to extract effective latent appearance representations and input them into recurrent networks. More crucially, we propose using attention models to detect salience in the spatial domain for each individual image as well as the temporal domain for the entire film. To extract the salient face regions in each individual image, we construct a specialized spatially-indexed attention mechanism among the convolutional layers, as well as a temporal attention layer to give attention weights to each frame. This two-pronged strategy not only increases performance by allowing the model to focus on informative frames and facial areas, but it also provides an interpretable relationship between the spatial and temporal facial regions, as well as the age estimation job. In studies on a large, gender-balanced database with 400 people ranging in age from 8 to 76 years, we show that our model performs well. Experiments show that, given enough training data, our model outperforms state-of-the-art approaches by a significant margin.