Person annotation in video sequences

Abstract

In the recent years, the demand for video tools to automatically annotate and classify large audiovisual datasets has increased considerably. One specific task in this field applies to TV broadcast videos, to determine who and when a person appears in a video sequence. This work starts from the base of the ALBAYZIN evaluation series presented in the IberSPEECH-RTVE 2018 in Barcelona, and the purpose of this thesis is trying to improve the results obtained and compare the different face detection and tracking methods. We will evaluate the performance of classic face detection techniques and other techniques based on machine learning on a closed dataset of 34 known people. The rest of characters on the audiovisual document will be labelled as "unknown". We will work with small videos and images of each known character to build his/her model and finally, evaluate the performance of the ALBAYZIN algorithm over a 2h video called "La noche en 24H" whose format is like a news program. We will analyze the results and the type of errors and scenarios we encountered as well as the solutions we propose for each of them if there is any. In this work, We will only focus on a monomodal basis of face recognition and tracking.