MacDorman, K., Green, R., Ho, C. and Koch, C. (2009). Too real for comfort? Uncanny responses to computer generated faces. Computers in Human Behavior, 25(3), pp.695-710.

The reason why I chose this article is because, as a character artist, the Uncanny Valley is a phenomenon that has, for many years remained a limiting factor in the creation of realistic and believable human faces. This article was initially published in the Computers in Human Behavior Journal in 2009 by Robert D. Green, Chin-Chang Ho, Clinton T. Koch and Karl F. MacDorman who is currently an associate professor in the School of Informatics and Computing in Indiana University, working on the Human-Computer Interaction program (MacDorman, 2016). The journal itself focuses on the use of computers from a psychological perspective (Behavior, 2016). This I find to be a particularly interesting perspective as it focuses on the psychological side of the Uncanny Valley, which despite major innovations in GCI character rendering remains unconquered.

As mentioned in the article, the uncanny valley effect is primarily caused by the increasing levels of photorealistic perfection which in turn make audiences feel more uncomfortable, (Karl F. MacDorman et al., 2009) when compared with more stylized faces. This paper relates to my work as it explores psychological theories like empathy, mate selection, threat avoidance, cognitive dissonance and psychological defenses when it comes to character perception and looks at how individual factors like facial proportions, skin texture and level of detail can increase the believably of a character’s face and can in turn directly aid me in the creation of more believable or unsettling characters.

The paper starts with a breakdown of the aspects that the researchers suspect are the key contributors to the uncanny valley effect when it comes to human characters as well as an overview of the intended methodology. In this paper the research is rendered in 4 separate studies: Study 1 – Baseline eeriness and human likeness with varied levels of detail, Study 2 – Sensitivity to best proportion the effect of warped facial proportions on faces with different detail levels, Study 3 – Eeriest level of detail which aims to find which proportions when pushed to the limit can be perceived the eeriest at high levels of details and Study 4 – Eyes–face mismatch which further studies how mismatch in various facial elements can make them seem eerie. In the end the paper synthesizes several basic design principles for CG animators and character artists that advise to combine lower quality textures with a stylized character model or if dealing with realistic textures, focusing on ideal facial proportions.

For this evaluation I will focus on the final two pages of the paper where the authors focus on summarizing and defining their findings from the four studies mentioned earlier. An overall observation of the results led the authors to conclude that faces which were regarded as looking more human had higher levels of detail in terms of model geometry as well as using photorealistic textures. They continue by saying that photorealistic textures on a human face made it easier for subjects to identify the model as natural looking or opposed to eerie looking when combined with varied facial proportions. However Study 4 showed that the most natural or realistic CG face was not perceived as the least eerie, instead participants agreed that the most natural or believable looking GC face was one that relied on only 75% on photorealism, while being combined and bronze and line texture materials. Claims such as these can be difficult to accurately measure or accredit to a specific feature, but the authors have largely managed to isolate and link outcomes to variables by isolating specific image aspect like materials, models and textures in separate and combined experiments mentioned earlier. However much of the data collected here should be taken with a dose of suspicion. This is due to the fact that studies on the uncanny valley effect (as mentioned under limitations) generally lack a common unit of measure. While the eeriness scale used for this study does unify the results of the different studies made here, there is no coherent method or unit used to compare the outcome of this paper with other similar ones like ”To Stylize or not to Stylize? The Effect of Shape and Material Stylization on the Perception of Computer-Generated Faces” (Zell et al., 2015) which picks up where this paper (Karl F. MacDorman et al., 2009) concluded – attempting fine tune the stylization and realism ratio mentioned in the conclusion here.

Another limitation to the studies is the level of control in the form of the base model as the only variations used in the study are that of a white middle aged male this can be seen as a limiting factor to some of the recommendations as stylization can have many varied and unpredictable effects on characters of different ethnicity or gender which are not explored in this study.

Another field I feel was left unexplored is the rendering aspect of a face. The method used to display the final CG image can have a drastic effect on the overall pre4sentation and believably of the model. As such there are currently several methods including phong shading, matcap materials and physically based render shaders, (the last of which is the most accurate when it comes to rendering objects in a lifelike manner) which I believe must be addressed when mentioning photorealistic rendering. Expanding on that I think that one major limitation of the study especially in terms of looking at photorealistic textures is not using a physically based shader material for the face. This means that in terms of photorealistic textures there is no mention of bump, specular, roughness or subdermal maps – all of which play a key part in showing detail (polycount as seen in this study is somewhat irrelevant), smoothness or reflectivity and control of subsurface scattering, which are what makes CG skin believable.

Overall while this paper does overlook some technical factors regarding rendering or base model variation, it manages to synthesize some essential guidance for character creation that is based in psychological theory. I believe that the findings here are important as they show several correlations between texture, model and material quality and how different quality levels can produce results of varied believably. A key lesson from this paper is the fact that while photorealism can take a design up to the uncanny valley it is through stylization and artistic liberty that it can cross it.

 

Bibliography

Behavior, C. (2016). Computers in Human Behavior – Journal – Elsevier. [online] Journals.elsevier.com. Available at: http://www.journals.elsevier.com/computers-in-human-behavior [Accessed 18 Dec. 2016].

MacDorman, K. (2016). Karl F. MacDorman. [online] Macdorman.com. Available at: http://www.macdorman.com/ [Accessed 10 Dec. 2016].

MacDorman, K., Green, R., Ho, C. and Koch, C. (2009). Too real for comfort? Uncanny responses to computer generated faces. Computers in Human Behavior, 25(3), pp.695-710.

Zell, E., Aliaga, C., Jarabo, A., Zibrek, K., Gutierrez, D., McDonnell, R. and Botsch, M. (2015). To stylize or not to stylize?. ACM Transactions on Graphics, 34(6), pp.1-12.

Advertisements