Single-Modal Video Analysis of Personality Traits using Low-Level Visual Features

The ability to analyze the first impression of a person automatically enables novel applications in human-computer interaction and other areas. A person's first impression can decide about a positive or negative outcome in different daily-life situations. The human brain is able to get a picture of the counterpart's personality at short notice. The main aim of this paper is to show how a system based on a standard Convolutional Neural Network (CNN) as well as a 3D-CNN architecture, can be built to solve a multi-label regression task using only visual low-level features. This paper investigates how various pre-processing methods, such as face-extraction and data-augmentation, influence the predicted personality confidences. Furthermore, it explores different training strategies and optimization techniques e.g. regularization in order to improve the model performance. The results of this paper demonstrate an image-based as well as a time-sequence-based system to predict the Big-Five personality dimensions of humans in short video sequences by solving a multi-label regression task. The approaches reach an overall accuracy of 0.891 by using visual features only.