ADVISE: Advanced Digital Video Information Segmentation Engine Presented by Ng Chung Wing
Jan 02, 2016
ADVISE: Advanced Digital Video Information
Segmentation Engine
Presented by Ng Chung Wing
OutlineIntroductionOverview of ADVISE System architecture and services provided
Technologies in ADVISE Construction of Video Table-of-Contents (V-To
C) Video Summarization Video Matching
Conclusion
Introduction
IntroductionVideos is getting more popular in education, entertainment and information sharingEvident growth of video contents on the Internet
57.2% of Internet users watched video chips and 7.3% edited video clips on their personal computers (Survey by PC Data, 2000)
Important task retrieve an interested video!Two problems in video retrieval: Not enough information to describe the video
contents Difficult to search for videos with similar
contents
Introduction
IntroductionWe propose the “ADVISE” system to solve the above problemsADVISE Advanced Digital Video Information
Segmentation Engine Web-based video browsing and retrieval
system Provides a set of services:
For description of videos: Video table-of-contents (VToC) Video summarization
For searching similar videos: Video feature similarity matching
Introduction
ContributionsWe propose the framework of a video browsing and retrieval system called ADVISEWe build the image-based video description called Video Table-of-Contents (V-ToC) in ADVISEWe develop the Video Summarization Algorithm for generating video summaries in ADVISEWe propose two Video Tree Matching Algorithms, which measure the feature similarity between videos, in ADVISE
Overview of ADVISE - OutlineIntroduction
Overview of ADVISE Objective of ADVISE System Architecture Services provided by ADVISE
Technologies in ADVISEConclusion
Overview of ADVISE
Objectives of ADVISETo provide an efficient way to describe the video contentsTo save the time for browsing the whole video to know the contentsTo search videos with similarity in certain video featuresTo provide services through the Internet
Overview of ADVISE
Provide services to users of ADVISE
Major contributions
System Architecture of ADVISE
ADVISE consists of 3 modules
(I) Video preprocessing module
(II) Web-based retrieval module
(III) Video streaming server
Process source videos
Setup the Real System Server for delivering video contents
Setup the Real System Server for delivering video contents
(III) Video streaming server
Overview of ADVISE
System Architecture of ADVISE
Input Raw Video
Web-based Video RetrievalModule
Web-based Video RetrievalModule
Internet User Terminal
1. Query for theV-ToC to know the
video contents
2b. Return CustomizedSMIL Video Summary
to user
2a. SubmitSelection Request
XML V-TOCspresented with
XSL
Video PreprocessingModule
Video Matching
V-ToC Construction
Video MatchingResultsStreaming Video Server
Streaming Video Server
Indexed VideoDatabase
Generation ofSMIL VideoSummary
3. Query onsimilar videos
ADVISE
III
III
Overview of ADVISE
Service 1: V-ToC Presentation
Service 2: Generation of SMIL Video Summary
Service 3: Querying Similar Videos
(II) Web-based Video Retrieval Module
User interface for accessing services of ADVISEThis module reside on a web serverThere are 3 services provided by ADVISE
Web-based Video RetrievalModule
Web-based Video RetrievalModule
Internet User Terminal
1. Query for theV-ToC to know the
video contents
2b. Return CustomizedSMIL Video Summary
to user
2a. SubmitSelection Request
XML V-TOCspresented with
XSL
Video MatchingResults
Generation ofSMIL VideoSummary
3. Query onsimilar videos
1
2
3
Overview of ADVISE
Module (II) ~Service 1: V-ToC Presentation
Image-based description for the video content Use the V-ToC structure resulting from Module (I)
Used XML with XSL to generate a flexible web-based presentation
Each V-ToC show us the contents organization in a video
Overview of ADVISE
Module (II) ~ Service 2: Generation of SMIL Video Summary
Generate a video summary according to user’s preferenceUsed SMIL to deliver the customized video summary to the user
User’s input
Resulting SMIL Video Summary
Overview of ADVISE
Module (II) ~Service 3: Querying Similar Videos
Show similar videos in descending order of the similarity score Results of video matching in Module (I)
User can select matching different video features Color histogram feature Shot style feature
List of videos with differentsimilarity scores with the queryvideo
Technologies in ADVISE - OutlineIntroductionOverview of ADVISE
Technologies in ADVISE For service 1: Construction of Video Table-of-Conte
nts (V-ToC) For service 2: Video Summarization For service 3: Video Matching
Conclusion
Technologies in ADVISE - Construction of V-ToC
Construction of Video Table-of-Contents (V-ToC)
Uses Image-based video description which show
the organization of video contents Video Table-of-Contents (V-ToC)
Provide the hierarchy for structural matching of video
Video structure used in ADVISE Hierarchical tree structure with 4 levels
Storage and Presentation Use XML and XSL
Technologies in ADVISE - Construction of V-ToC
Video Structure in ADVISEDecompose a video into 5 levels: Video Frames Video Shots Video Groups Video Scenes Whole Video
Hierarchical Representation of a Video
Technologies in ADVISE - Construction of V-ToC
Video Structure in ADVISEExample:
Group 1
Group 3
Group 2
Scene 1
Scene 2
Video
Shots:
Shot 1
Shots 2,4,6
Shots 3,5,7
12
3
4
Technologies in ADVISE - Construction of V-ToC
Video Structure in ADVISEStructure videos from the bottom level5 steps in video structuring
i. Color Histograms Extraction ii. Video Shot Boundaries Detection iii. Video Groups Formation iv. Video Scenes Formation
Vid
eoS
hot 4
Vid
eoS
hot 6 ...
Vid
eoS
hot 2(Video Shot Level of V-ToC)
...(Video Group Level of V-ToC)
Vid
eoS
hot 3
Vid
eoS
hot 5
Vid
eoS
hot 7...
VideoGroup 1
VideoGroup 2
VideoGroup 3
VideoGroup 4
VideoGroup 5
Video Scene 2Video Scene 1 Video Scene 3 ...(Video Scene Level of V-ToC)
(Video Level of V-ToC) The Whole Video
Technologies in ADVISE - Construction of V-ToC
Storage and PresentationResulting Presentation of V-ToC using XML and XSL
Folding/Unfolding Video Components
Sequential Video Story Line
Enlarging Images
Basic Video Information
<?xml version="1.0"?><!DOCTYPE advise SYSTEM "./toc.dtd"><advise><video length ="25" src="rstp://localhost/video1.rm"><scene id="1"> <group id="1"> <shot id="1"> <keyframe img="./sh_1.jpg"/> <time value="0"/> </shot> <shot id="2"> <keyframe img="./sh_2.jpg"/> <time value="11"/> </shot> </group></scene></video></advise>
Technologies in ADVISE - OutlineIntroductionOverview of ADVISE
Technologies in ADVISE For service 1: Construction of Video Table-of-Conte
nts (V-ToC) For service 2: Video Summarization For service 3: Video Matching
Conclusion
Technologies in ADVISE - Video Summarization
Video SummarizationUser may still not be able to know the exact video contents with V-ToCVideo summary can provide all types of information in the videoObjectives:
Select the major contents Shorten the duration for browsing
Difficulties No standard method to pick the important contents from vi
deo Importance of contents depends on user’s need
In ADVISE: We accept user’s input for generating video summary suc
h that the result can be the best suitable for the user
Technologies in ADVISE - Video Summarization
Inputs for Video Summarization Algorithm
Video features used: Human faces Male and female voices Volume level Caption text
User’s inputs for customization of the video summary
Weights of different video features
Time constraint for video summary
Clustering control constant
Technologies in ADVISE - Video Summarization
Video Summarization Algorithm4 steps to summarize a video
i. Combining extracted video segments ii. Scoring the extracted video segments iii. Selecting extracted video segments iv. Refining the selection result
Example:
f1
Male Voice
Female Voice
Human Faces
Volume > 35dB
Caption Text
Feature Exists
Feature Not Exist
Video Frames Sequence
f80 f100 f300 f500 f5000 f5300 f6000
Step (i)
Step (iv)
Step (ii)
e7
e1 e2
e3e4
e6
f1 f80 f100 f300 f500 f5000 f5300 f6000
Video Frames Sequence
Sco
re fo
r V
ideo
Seg
men
ts
22
e5
21
6
5
3
0
Step (iii)
e7
e1
e3e4
e6
f1 f80 f100 f300 f500 f5000 f5300 f6000
Video Frames Sequence
22
e5
21
6
5
3
e2
Selected Segment
0
Sco
re fo
r V
ideo
Seg
men
ts
u1
Clu
ster
Sco
res
VideoFrames
u2 u3
u4
SelectedCluster
ea1
ea3
Video Frames Sequence
e a2
SelectedSegment
ea4
ea5 e a6
Short DisjointedSpaces
ea1
ea3
Video Frames Sequence
ea2
SelectedSegment
ea4
ea5 e
a6
Cluster u1
Cluster u2
Technologies in ADVISE - Video Summarization
Video Summary in SMILSMIL presentation are delivered to user of ADVISE Can be generated instantly Can be browsed by the user on the Internet
using a stream-based protocol
Resulting SMIL video summary
Technologies in ADVISE - OutlineIntroductionOverview of ADVISE
Technologies in ADVISE For service 1: Construction of Video Table-of-Conte
nts (V-ToC) For service 2: Video Summarization For service 3: Video Matching
Conclusion
Technologies in ADVISE - Video Matching
Video MatchingVideo Matching
Match the extracted video features Color, motion, shape, etc.
Sequential matching Non related to video structure
VToC is a tree structure Can apply tree matching algorithm Matching related to video structure
In ADVISE, we propose two tree matching algorithms (1) Non-ordered tree matching algorithm (2) Ordered tree matching algorithm (Consider temporal or
dering)
Technologies in ADVISE - Video Matching
Input Features for Video MatchingTwo video features used Color histograms feature
Take the first frame of a video shot as the key frame to compare in order to reduce the computational complexity.
Compare the visual similarity. Shot style feature
Compose of camera motion and length of a video shot.
Select the first camera motion in a video shot as the representative.
Compare the similarity in video pace.
Technologies in ADVISE - Video Matching
(1) Non-ordered Tree Matching Algorithm
Not constrained by temporal orderingCapture all similar components
Algorithm Extract features at the bottom level Propagate the similarity score up to the root
level
Group a1 Group b2
Video A Video B
Group a2 Group b1
Technologies in ADVISE - Video Matching
(2) Ordered Tree Matching AlgorithmConstrained by temporal orderingTemporal ordering can affect the video contents Reduce the problem to match components in
order
Capture only ordered similar components
Group a1 Group b2
Video A Video B
Group a2 Group b1
Technologies in ADVISE - Video Matching
(2) Ordered Tree Matching AlgorithmAlgorithm: Recursive dynamic programming
Hierarchical matching From the tree root (video level) Until shot level, extract the video feature
similarity
Reduced the complexity compare with approach (1)
sim(Videoa,Videob) =
0 0 0 0
0 D(2,2) D(3,2) D(4,2)
0 D(2,3) D(3,3) D(4,3)
c1
cb1
cb2
D
Videob
Videoa
c2 c3
D(4,3)
2
0 0 0
0 D1(2,2) D1(3,2)
0 D1(2,3) D1(3,3)
ChildSim(1,1)
=
g2
C1of Video b
D1g1
C 1of Video a
g1
g2
D1(3,3)
2
0 0 0
0 D2(2,2) D2(3,2)
0 D2(2,3) D2(3,3)
g1of Videoa
g1of Videob
D2 s1
s5
s1
s3
ChildSim( )
Conclusion
ConclusionThe ADVISE system, which enhanced video browsing and retrieval system on the Internet, is proposed.The generation and presentation of the image-based video description are developed.The automation of video summarization into SMIL format is provided.Two video tree matching algorithms for measuring the similarity between videos are proposed.
Questions & Answers