Aqua Phoenix
     >>  Research >>  Actor Interaction Vis  


2. Data

Actor interaction graphs are based on actor- and location-segmented and annotated information. Depending on the semantics and desired granularity of analysis, actor segmentation can be performed precisely by determining each actor's temporally disjunct talk units (e.g. through manual or automatic speaker segmentation), or more coarsely and with high-level human understanding by determining each actor's significant verbal contribution to a scene (i.e. disregard the occasional chuckle or short verbal interruption).

The data we are using for our experiments enumerates shots with actor and location annotations only for significant verbal or other dialogue contributions. Shots, in which primarily one actor is speaking, even though another actor may be visible but without a contribution to the dialogue, are recorded with the primary actor's ID only. Shots, in which two or more actors are involved in a dialogue with each other are marked as a list of those actors' IDs. A finer-granularity segmentation is omitted for these "group dialogue" shots.

Shots with actor IDs and other information are enumerated consecutively in rows in the shot spreadsheet. Empty breaks between rows are inserted when two shots are not contextually related. This occurs heuristically during scene breaks, which usually coincides with location changes. There are a few notable exceptions to location changes, when, for example, a dialogue takes place over the phone and the camera captures both parties in their respective locations. Two shots may also be contextually unrelated within a scene, when a dialogue interrupts or ends, and a new dialogue begins. For example, in a coffee shop, an actor may be part of a conversation with a friend while having coffee, then end the conversation, walk to the cashier, and begin a new conversation. The two dialogues are exclusive and are separated in the shot spreadsheet by a break.

Interaction is now defined as any actors appearing consecutively in the shot spreadsheet. When an actor in one shot follows an actor in the next shot, a unit is added to this actor pair's interaction "major" weight. When a group of actors appears in one shot, and one or more actors appear in the next, a unit is added to each combination of actor pairs between the two shots (also interaction "major" weight. In addition, when several actors are enumerated within a shot, an interaction "minor" weight is defined for each combination of actor pairs. Major and minor weights are displayed in the interaction graphs.

The shot spreadsheet used for this project enumerates shot boundaries, unique actor IDs, and location IDs, among several other criteria. The interactive tool presented here uses only scene markers, actor IDs, and location IDs to generate the display.

Figure 1.1: Shot Spreadsheet
Click image to enlarge, or click here to open
Figure 1.2: Shot Spreadsheet Header
Click image to enlarge, or click here to open