1 Shot

1.1 Introduction

The video shot annotations consist of two parts.

* Shot boundary: cut transition and gradual transition

* Shot type: Far-view shot, medium-view shot, close-view shot, out-of-field shot, playback shot. Figure 1 provides an example of each shot type.

1221D  ED77

(a)far-view shot                                                                (b)medium-view shot

B073  26962

(c)close-view shot                                                                  (d)out-of-filed shot

Figure 1: Shot type examples

A playback shot in soccer videos is usually sandwiched between two gradual transitions, which often contain the competition logo. The match logos vary with different competitions, and there are four logos in our dataset corresponding to the tournament in which the match is being played; examples are shown in Figure 2 below:



Figure 2: Logos present the dataset

“Other” -- a special label in shot boundary annotation in this dataset. Sometimes a single shot may contain more than one shot type; for example, a shot may start in medium view and then transition to far view. We have marked the boundary of these shots as "other", and annotated all shot types included during these shots, in order.

The annotation statistics for the different shot types and shot transitions are shown in Table 1:

Table 1: Shot annotation statistics

1.2 Shot Annotation files

Each shot is annotated with the first and last frame number and the shot transition type. There are two files relating to the shot annotations for each video:

* Shot boundary: (x, y)

Gradual transition :

(start-point, end-point)

Cut transition :

(frame no., -1)

Other :

(frame no., -2)

* Shot type. (start-point, end-point, shot type)

2 Event and story

We define two types of event boundary:

* Event: it is fine-grained with action-driven definition;

* Story: it has coarser granularity when compared with the boundaries of an event, and complete temporal context of an event.

2.1 Event

11 manually identified key event types are listed in Table 2. The boundaries of these events are defined according to the key player's action and the position of the ball. The "start point" and "end point" of an event are defined as:

* Start point: The event is at the beginning of the key player action (e.g., the player moves his legs to shoot). The "yellow-card" and "red-card" events start at the time that the yellow or red card are shown in the shot.

* End point: When the ball lands somewhere -- such as the goal or out of the field -- or is received by other players. The "yellow-card" or "red-card" event ends when the card disappears.

Besides of events beyond a replay segment, the replay segment for an event -- including the gradual transitions preceding and following the replay segment -- is also annotated with one of these event labels. Event examples are shown in Figure 3.

Table 2: Statistics of event and story


Figure 3: Examples of event and story annotations. The words in blue describe the state of the game. The words in orange are the event annotations, and those in green are the story annotations.

2.2 Story

The definition of story and event differ in two aspect just as shown in Figure 3:

* Boundary: Compared with the accurate localization for events, story annotations are based on the needs of audiences. A story includes the complete temporal context, with more interesting segments.

* Even type: The story types are similar to event types, other than the four labels for a union story shown in Table 2. Two or more consecutive events may occur in a story. We call these "union stories" and classify the consecutive events into a single story, such as "corner-shot". Another union story is "yellow-card" or "red-card". If a foul results in a card event, then the story -- including the foul event and yellow/red card event -- is annotated as a yellow-card or red-card, rather than a foul.

The definition of the "start point" and "end point" of a story are more complex than the definition of an event, as explained below:

* Start point: For a shot and goal, the story starts from the last delivery before the shot and goal. For set-pieces such as a free kick, corner, or penalty kick, the story begins when the player is at the designated position, and ready to kick the ball. The start-point of a foul or offside is the same as that of an event; i.e., at the beginning of a player's action.

* End point: The story ends at the end point of the replay for an event if the replay follows the event, otherwise when the ball lands, the ball is received by other players, or the foul ends.

2.3 Event/story Annotation files

There are two files relating to an event annotation -- one for the event and another for the story.

Event :

(event type, start-point, end-point, flag)

Flag = 1: it is a replay clip; otherwise flag is 0.

Story :

(story type, start-point, end-point)

3 Player tracking

* We focus on players in far-view shots, and the bounding box is scale-invariant.

* 19908 frames, including 40 shot sequences, were handpicked from soccer videos.

* It contains 80 ground truths for different object players (there is at least one player chosen as object in each shot sequence).

Those shots covering the four sets of circumstances shown in Figure 4. And each set of circumstances contains 20 tracking samples. The four sets of circumstances are:

* General circumstance. The player's posture changes without a block or any other interference.

* The region of object player overlap with that of another team-mate.

* The player region overlaps with that of a player in another team.

* The complex circumstance. Players are concentrated, and the player region overlaps with different players.









Figure 4:  Four short player tracking sequences corresponding to four sets of circumstances

3.1 Player tracking Annotation files

* First row of the ground-truth file: (start frame no., end frame no.) represents the frame id where the player tracking starts and ends in this shot sequence.

* Each row in the rest represents the bounding box of the object in that frame, (x, y, box-width, and box-height).

 Laboratory of Digital Media Processing and Retrieval,HUST.  © 2017. All rights reserved.