1. Football Video Segmentation Based on Video Production Strategy
Authors: Reede Ren and Joemon M. Jose
Conference paper 2005
Noureddine Ghoggali
12/9/2018
Noureddine Ghoggali

2. Attack structure definition Attack structure detection Data set
Outline
Introduction
Goal of the paper
Methodology
Attack structure definition
Attack structure detection
Data set
Application
Conclusions
12/9/2018
Noureddine Ghoggali

3. Introduction
The amount of audiovisual information available in digital format has grown exponentially in recent years. Gigabytes of new images, audio and video clips are generated and stored everyday. This has led to a huge distributed and mostly unstructured repository of multimedia information. In order to realize the full potential of these databases, tools for automated indexing and intelligent search engines are urgently needed. Indeed, image and video cataloguing, indexing and retrieval are the subject of active research in industry and academia across the world. This reflects the commercial importance of such technology. The aim of this presentation is to give an overview of Football video segmentation.
12/9/2018
Noureddine Ghoggali

4. Basic concepts
Cognitively, the predominant feature in video is its higher-level temporal structure. People are unable to perceive millions of individual frames, but they can perceive episodes, scenes, and moving objects. A scene in a video is a sequence of frames that are considered to be semantically consistent. Scene changes therefore demarcate changes in semantic context. Segmenting a video into its constituent scenes permits it to be accessed in terms of meaningful units.
A video is physically formed by shots and semantically described by scenes.
shot is a sequence of frames representing continuous action in time and space.
scene is a story unit and consists of a sequence of connected or unconnected shots.
12/9/2018
Noureddine Ghoggali

5. Basic concepts
Most of the current research efforts are devoted to shot-based video segmentation. Algorithms for scene change detection can be classified according to the features used for processing into uncompressed and compressed domain algorithms. Temporal segmentation is the process of decomposing video streams into these syntactic elements. Shots are a sequence of frames recorded continuously by one camera and scenes are composed of a small number of interrelated shots that are unified by a given event.
12/9/2018
Noureddine Ghoggali

6. Video production strategy
Video Production Strategy in Football Broadcasting
A football game is made up by a series of team movements called attack in sports jargon. They are mostly independent and sorted by time throughout the game. In some sense, attack decides broadcasting strategy.
During broad-casting, video reporters focus on two issues,
how to record the game or attack
how to avoid missing interesting issues in an attack.
They employed view to describe team tactics and middle view or close-up view to catch players detailed movement. When an important event or highlight takes place, such as goal, it will be replayed.
12/9/2018
Noureddine Ghoggali

7. Video production strategy
The strategy can be stated as following,
When an attack begins, a global view will be used until the ball passes the centre circle.
When the ball comes into front field, a middle view is going to be employed to show how groups of players attack and defend.
When the ball come into or close to the penalty area, a close-up view is here to catch possible highlights and players action in detail.
When there is a highlight, such as shoot and foul, a close-up slow motion replay will come to state the event.
12/9/2018
Noureddine Ghoggali

8. Video production strategy
With these observations, we conjecture:
Video making methods in football game dictate the structure of video and compose semantics.
As a time sequence, a football game can be modeled by Hidden Markov Model with attack video structure.
Attack is an independent semantic video unit, which can be treated as a scene in football video domain.
12/9/2018
Noureddine Ghoggali

9. Goal of paper
Parse continuous video stream into a sequence of ‘attack’.
Set up a hierarchical video content index to summarize the game and allow a non linear navigation of video content.
12/9/2018
Noureddine Ghoggali

10. Methodology
12/9/2018
Noureddine Ghoggali

11. Attack definition
Attack takes the role of scene in the framework. To detect it, the authors define a new video structure layer between shot and attack It includes four mutually exclusive video structures in broadcasting video data (play, focus, replay and break).
Play: video makers convey global status of the game and employ
long and medium shots or field view .
Focus: short stop of game, in which the video maker traces a player to show his or her detailed actions.
Replay: is for slow-replays.
Break : includes non-game video clips.
12/9/2018
Noureddine Ghoggali

12. Attack structure detection
A two pass classification is proposed to identifies video structures and labels video frames with their video structure type: play, focus, replay and break by a GMM classifier and its output is smoothed by dynamic programming process
Second pass detects replay. Replay is a slow motion video clip sandwiched by editing effects.
12/9/2018
Noureddine Ghoggali

13. Attack structure detection
Three salient features are computed for the first pass GMM classification to parse video in attack structure
Field ratio: play field area ratio over image.
Zoom size: to detect football uniform.
Edit effect detection: detecting logos.
12/9/2018
Noureddine Ghoggali

14. Attack structure detection
12/9/2018
Noureddine Ghoggali

15. Attack structure detection
Field ratio:
 its reasonable to expect grass color to be the dominent color.
12/9/2018
Noureddine Ghoggali

16. Attack structure detection
Zoom size:
Football uniform is an oblivious domain feature compared with human face. It has the following merits:
its with bright color and special pattern
it associates with the appearance of player only.
It is rotation robust.
12/9/2018
Noureddine Ghoggali

17. Attack structure detection
Zoom size: use the Foley-Sammon Transform algorithm to build the training set
Edit effect detection:
12/9/2018
Noureddine Ghoggali

18. Attack scene construction
After attack structure classification, we get the video structure label sequence "...BPFPFPRP...", where B is the abridgement for `Break', P for `Play', F for `Focus' and R for `Replay'. The string records the process of video making and keeps the information of `attack'.
So the job of `attack' construction is to divide it into a serials of substrings, which contain only one `attack' sequence each.
But the string is too long for the attack.
12/9/2018
Noureddine Ghoggali

19. Attack scene construction
12/9/2018
Noureddine Ghoggali

20. Experiments
Data Set:
The data set includes two MPEG-1 broadcasting videos in World Cup 2002, the final game and the one Japan vs Turkey. It is about 320 minutes, containing interview, celebration and commercial clips. Both games are divided into halves, Final I, Final II, Japan-Turkey I and Japan-Turkey II.
The first half of Japan-Turkey and final game are labelled manually to set up ground truth.
4535 play frames (33.7%),
4253 focus frames(31.6%) and 4674 `break' frames (34.6%).
There are 33 replay in the final game and 34 in the Japan-Turkey game.
Training set:
includes 2000 frames
400 from play
1000 from focus, 600 from break, which are randomly selected from marked samples. Remaining frames are kept for test.
The grass hue model is automatically calculated for every game.
12/9/2018
Noureddine Ghoggali

21. Experiments
12/9/2018
Noureddine Ghoggali

22. Application
Related video browser (a) retrieves replay' and its `play' and `focus' segments. It includes two regions in the panel, replay segment list and related segments panel. The related segments panel displays top n (n=3) closest `play' and `focus‘ to the selected `replay'. User chooses `replay' from the `replay' segment list and decides whether to contain it and its related video segments in summary or not. A double-click on icons will play the video clip by a stand alone window.
Summary browser (b) shows all video segments in the proposed summary, and grantees user the ability to insert and remove shots.
12/9/2018
Noureddine Ghoggali

23. Conclusions
football videos. It is based on video production conventions and helps in video summarization and indexing.
attack is a semantic unit of football game and is an equivalent of scene in other video domains.
The result shows those high-level video structures can be computed with high accuracy.
We focus on video structure identification and how to merge these
structures into attack scene.
future work:
measure accuracy of attack boundary. The algorithm leaves much space for improvements:
Audio event detectors, such as goal and whistle detection, can be integrated;
Improve GST algorithm to search more embedded video structure;
12/9/2018
Noureddine Ghoggali

24. Comments
The major difficulty with the statistical approaches in that a large number of hand labeled training samples are required to learn correctly. This problem becoming crucial. And thus the expert are required to derive the semantic labels for a large number of training samples. Given this cost problem its very attractive to design more effective classifier.
Training techniques that can take advantage of the unlabeled samples
There are three primary genetic operators that govern the producing of the new offspring in the next generation
12/9/2018
Noureddine Ghoggali

25. THANK’S 12/9/2018 Noureddine Ghoggali
There are three primary genetic operators that govern the producing of the new offspring in the next generation
12/9/2018
Noureddine Ghoggali