Face Analysis Under Uncontrolled Conditions

Conclusion to Part 1

Romain BELMONTE¹, Pierre TIRILLY¹, Ioan Marius BILASCO¹, Nacim IHADDADENE² and Chaabane DJERABA¹

¹University of Lille, France

²Junia ISEN, Lille, France

In the previous chapters, we have investigated facial landmark detection under uncontrolled conditions. The first line of work focused on performance analysis of current methods. We highlighted a weakness in commonly used evaluation protocols: they rely on datasets with limited annotations and do not allow for exhaustive performance analysis. It is especially complex to quantify the performance of approaches according to the challenges encountered. Our efforts were therefore directed towards this issue.

The second line of work focused on the development of a solution capable of coping with most of the difficulties encountered under uncontrolled conditions (e.g. pose, expression, illumination, occlusion, motion blur). Based on the results of our initial work, we considered video-based solutions. Unlike the popular still-image approaches which are unable to take advantage of temporal information, video-based approaches leverage temporal coherence to provide more robustness under uncontrolled conditions. We noted that the approaches proposed so far provide only partial temporal modeling and addressed this issue.

C1.1. Summary

In this part of the book, contributions for facial landmark detection are presented in both the analysis of the performance of current approaches and in the modeling of temporal information. ...