One Ear and Two Eyes: Effects of camera angle on character recognition in film.

Camera angle of a character

Camera angles are an important part of composition in film, photography and animation. Both vertical angle and the horizontal angle can be changed to convey different emotions and meaning. A high vertical angle shot can make the subject appear physically diminished and insignificant, and a low angle shot can make the subject appear more important and dominant in the frame. Altering the horizontal angle changes apparent depth of the face and the visible facial features.

In this study we focus on horizontal camera angles and the effects of camera angle on character recognition in film. We conduct an online study to test our hypothesis, by showing the participants characters faces from five different angles and measuring the accuracy of recognizing them in a short film segment.

Related Work

There are several studies that explore facial recognition and facial perception. A 2007study by Benton at el on the viewpoint dependence of facial expression after-effects suggest that there remains a substantial after-effect when adapt and test are at differing three-quarter views [1]. A 2010 study by Berisha, Jonston and McOwan on Identifying regions that carry the best information about global facial configurations concluded that regions of the face containing the greatest identifying power include the hairline, facial shape, eyes, and mouth [2]. Studies have also shown that more attractive faces (and, likewise, notably unattractive faces) are among the easiest to recognize [9].

See Faces in fim for more related work.

Methodology

Experimental Design

We conducted a 5 (Camera Angle: left, mid-left, center, mid-right, right) X 5 (Character: Alex, Brandon, Dan, Eric, Frank) multi factorial mixed model experiment. The factors Camera angle and Character were within participant variables, and combination Camera Angle X Character was between participant variables.

Camera Angle

The Angle the character is introduced to the participants in the priming stage

  1. left: A side shot where left ear of the character is visible and left eye maybe partially visible
  2. mid-left: An intermediate shot where left ear and both eyes of the character is visible
  3. center: A straight shot where both eyes and both ears of the character is visible
  4. mid-right: An intermediate shot where right ear and both eyes of the character is visible
  5. right: A side shot where right ear of the character is visible and right eye maybe partially visible

Character

The Characters were extracted from the television series The Event and the faces were cropped out minimizing other identifiable artifacts for the priming steps. All the characters are adult white males and non of them are well known actors, and the names were assigned without any special consideration (Alex, Brandon, Dan, Eric, Frank). A sixth dummy character (Cindy) was introduced to balance the number of characters seen before each clip.

We acknowledge that the resolution, relative facial size, and facial expressions are not consistent between the characters and camera angles.

Hypothesis

We frame our hypothesis based on the camera angles that can be grouped together due to symmetry

  1. The character recognition will be most accurate when the priming is done from mid-left or mid-right camera angles.
  2. The character recognition will be least accurate when the priming is done from sides, left or right camera angles.

Experimental Procedure

To Investigate our hypothesis we utilized an online survey tool (Qualtrics), upon starting the survey the participants were instructed that they will be shown a group of screen captures and two short video clips and that they will be asked some questions after viewing each of the video clips. The first part of the task included looking at three character faces with the name of the character, a number and radio button displayed on the screen for each character, the participants have to select the radio button and click next to proceed. This was the priming stage and order in which the characters appeared was randomized and each character was displayed from one of the Camera angles.

After viewing the characters the participants were shown a video clip which contained two of the viewed characters among many other characters. Following the clip the participants were asked identify Characters they recognized in the video clip by name. They were also asked two dummy questions (eg: What number was displayed next to Alex) to prevent them from only focusing on the characters. The task was repeated again with 3 different characters and a different video clip.

Participants

Participants were recruited from three different sources for the online survey, which include UW-Madison grad chat, Facebook and Amazon Mechanical Turk. The full sample consisted of 23 males and 16 females with ages ranging from 18-52 (M = 29.75, SD = 8.01). The experiment was run as pre-test where the Mechanical Turk participants were paid $ 0.50 and the others participating voluntarily.

Measurements

For each of the 5 camera angles and each of the 5 characters we measure the accuracy in which they were recognized in the video segment as a single objective dependent variable. An accurate recognition would be

  1. The character was present in the video segment and was marked being present
  2. The character was not present in the video segment and was marked being absent

Results

We conducted a two-way analysis of variance (ANOVA) to test whether the camera angle and the character affected participants’ character recognition accuracy found a significant effect of character on recognition F(4,140) = 3.51, p < .01. Post-hoc comparisons using Tukey’s HSD test revealed differences between character A to B and E to be (p < .01) .

experiment results

Discussion

The results in this experiment did not provide significant proof to support our hypothesis, In fact the trend of the results although not statistically significant seem to contradict our hypothesis that the mid camera angles would result in best recognition accuracy. This could be a interesting factor to look in to in future work as it does not agree with the existing work on facial perception.

The fact the character A had significantly lower recognition accuracy is interesting as the character A’s priming images were of better resolution than some of the other characters. As mentioned previously we believe that the quality and consistency of characters images and video clips be improved and may yield to more convincing results.