TEMPORAL VIDEO BOUNDARIES -PART ONE- SNUEE KIM KYUNGMIN
Dec 26, 2015
Why do we need temporal segmen-tation of
videos?
How do we set up boundaries in be-tween video
frames?
How do we merge two separate but uniform
segments?
ABSTRACT
Much work has been done in automaticvideo analysis. But while techniques likelocal video segmentation, object detectionand genre classification have beendeveloped, little work has been done onretrieving overall structural properties of
avideo content.
ABSTRACT(2)
Retrieving overall structure in a video con-tent
means splitting the video into meaningful tokens
by setting boundaries within the video. =>Temporal Video Boundary Segmentation
We define these boundaries into 3 cate-gories : micro-, macro-, mega- boundaries.
ABSTRACT(3)
Our goal is to have a system for au-tomatic video
analysis, which should eventually work for
applications where a complete metadata is
unavailable.
INTRODUCTION
What’s going on?Great increase in quantity of video con-tents.
More demand for content-aware apps.Still the majority of video contents have insufficient metadata.
=> More demand for information on temporal video boundaries.
BOUNDARIES : DEFINITIONS
Micro-bound-aries : the shortest ob-servable tem-poral seg-ments. Usually bounded within a se-quence of con-tiguously shot video frames.
(frames under the same micro-bound-aries.)
Micro-boundaries are associated to the smallest video units, for which a given attribute is constant or slowly varying. The attribute can be visual, sound or text.
Depending on which attribute, mi-cro-boundaries can differ.
BOUNDARIES : DEFINITIONS(2)
Macro-bound-aries : boundaries between dif-ferent parts of the narra-tive or the segments of a video con-tent.
(frames under the same macro-bound-aries.)
Macro-boundaries are boundaries between micro-boundaries that are clearly identifiable organic parts of an event defining a struc-tural or thematic unit.
BOUNDARIES : DEFINITIONS(3)
Mega-Boundaries :
a boundarybetween a program and
anynon-programmaterial.
(frames under different mega-bound-aries.)
Mega-Boundaries are boundaries between macro-boundaries which typically exhibit a structural and feature consistency.
BOUNDARIES : FORMAL DEFINITION
A video content contains three types of
modalities : visual, audio, textual
and each modality has three levels : low-, mid,
high-
These levels describe the “amount of details”
in each modality in terms of granu-larity and
abstraction.
BOUNDARIES : FORMAL DEFINITION(2)
For each modality and levels is an attribute. An
attribute defined as below. (at-tribute vector)
: denotes modality( ex : m=1, 2 and 3 means visual, au-dio and text respectively.
: denotes the index for the attributes. (ex : m=1 and =1 indexes color )
: denotes the total number of vector components.
: time constant ( can be expressed in integers or milliseconds.)
BOUNDARIES : FORMAL DEFINITION(3)
If time interval is defined as , the average and
the deviation of an attribute throughout the
video can be expressed as below := avg of
(deviation) =
Where
BOUNDARIES : FORMAL DEFINITION(4)
By using the vectors defined previously, we now have
two different methods to estimate temporal boundaries :
Local Method Global Method
Has no memory Has memory
Given a threshold , and distance metric ‘Dist’, ifDist( )is larger than , then there exists a boundary at instant
The difference computed over a series of time. So we calculate the distance metric between the universal average, instead of the previous attribute.
If Dist ,a boundary exists at in-stant
MICRO-BOUNDARIES
In multi-media, the term “shot” or “take” is widely used.
Similar concept can be used to define the segment
between micro-boundaries, which is often called a
“family of frames.”
Each segment has an representative frame called
“keyframe.” The keyframe of a family has audio/video
data that well represents the segment. But the method
to pick out the keyframe may vary.
MICRO-BOUNDARIES(2)Each family has a “family histogram” to eventu-
ally form a
“superhistogram.”
A family histogram is a data structure that repre-sents
the color information of a family of frames.
A superhistogram is a data structure that contains the
information about non-contiguous family histograms
within the larger video segment.
MICRO-BOUNDARIES(3)
Generation of family histograms and superhis-tograms
may vary depending on pre-defined dimen-sions below.
1) The amount of memory
-No memory means comparing only with the pre-
vious frame.
2) Contiguity of compared families
-Determining the time step.
3) Representation for a family
-How we choose the keyframe.
MICRO-BOUNDARIES : FAMILY OF FRAMESAn image histogram is a vector representing
the color values and the frequency of their occurrence in the image.
Finding the difference between consecutive his-tograms and merging similar histograms en-able generating family of frames.
For each frame, we compute the histogram( ) and then search the previ-ously computed family histograms( ) to find the closest match.
MICRO-BOUNDARIES : FAMILY OF FRAMES(2)Several ways to generate histogram
difference :
Among them, the L1 and bin-wise histogram intersection gave the best results.
MICRO-BOUNDARIES : BOUNDARY DE-TECTIONIf the difference between two family his-
tograms is less than a given threshold, the current histogram is merged into the family histogram.
Each family histogram consists of :
1) pointers to each of the constituent histograms and frame numbers.
2) a merged family histogram.
MICRO-BOUNDARIES : BOUNDARY DE-TECTION(2)Merging of family histograms is per-
formed as below:
(basically, the mean of all histograms in the given video.)
MICRO-BOUNDARIES : BOUNDARY DE-TECTION(3)Multiple ways to compare and merge families,
depends on the choice of contiguity and memory.
1) Contiguous with zero memory
2) Contiguous with limited memory
3) Non-contiguous with unlimited memory
4) Hybrid : first a new frame histogram is compared using the contiguous frames and then the generated family histograms are merged using the non-contiguous case.
MICRO-BOUNDARIES : EXPERIMENTS
CNN News Sample.
27,000 frames
Tested with 9, 30, 90, 300 bins in HSB, 512 bins in RGB
Multiple histogram comparisons: L1, L2, bin-wise inter-section and his-togram intersec-tion.
Tried on 100 thresh-old values.
MICRO-BOUNDARIES : EXPERIMENTS(2)
Tested on a video clip, best results showed when threshold 10 with the L1 comparison/contiguous with limited memory boundary method/HSB space quantized to 9 bins.
MACRO-BOUNDARIES
A story is a complete narrative structure, con-veying a continuous thought or event. We want micro-segments with the same story to be in the same macro-segment.
Usually we need textual cues(transcripts) for setting such boundaries, but this paper sug-gests methodologies that does the job solely with audio and visual cues.
We focus on the observation that stories are characterized by multiple constant or slowly varying multimedia attributes.
MACRO-BOUNDARIES(2)
Two types of uniform segment detection :
Unimodal and multimodal
Unimodal(under the same modality) : when a video segment exhibits the “same” charac-teristic over a period of time using a single type of modality.
Multimodal : vice versa
MACRO-BOUNDARIES : SINGLE MODAL-ITY SEGMENTATIONIn case of audio-based segmentation:
1) Partition a continuous audio stream into non-overlapping segments.
2) Classify the segments using low-level au-dio features like bandwidth.
3) Divide the audio signal into portions of dif-ferent classes.(speech, music, noise etc.)
MACRO-BOUNDARIES : SINGLE MODAL-ITY SEGMENTATION(2)In case of textual-based segmentation :
1) If transcript doesn’t exist, extract text data from the audio stream using speech-to-text conversion.
2) The transcript segmented with respect to a predefined topic list.
3) A frequency-of-word-occurrence metric is used to compare incoming stories with the pro-files of manually pre-categorized stories.
MACRO-BOUNDARIES : MULTIMODAL SEG-MENTSWhat we want to do : Retrieve better seg-
mentation results by using the results from various unimodal segmentations.
What we need to do : first the pre-merging steps, and then the descent steps.
MACRO-BOUNDARIES : MULTIMODAL SEG-MENTS(2)
Pre-merging Steps : detect micro-segments that exhibit uniform properties, and deter-mine attribute templates for further segmen-tation.
1) Uniform segment detection
2) Intra-modal segment clustering
3) Attribute template determination -attribute template : a combination of numbers that characterize the attribute.
4) Dominant attribute determination
5) Template application
MACRO-BOUNDARIES : MULTIMODAL SEG-MENTS(3)
Descent Methods : By making combinations of multimedia segments across multiple modalities, each attribute with its segments of uniform values is associated with a line.
MACRO-BOUNDARIES : MULTIMODAL SEG-MENTS(4)
Single descent method describes the process of generating story segments by combining these segments.
1) Single descent with intersecting union
2) Single descent with intersection
3) Single descent with secondary attribute
4) Single descent with conditional union
MACRO-BOUNDARIES : EXPERIMENTS
Single descent process with conditional union.
Used text transcript as the dominant attribute.
-uniform visual/audio segments
-uniform audio segments
You can find a lag be-tween the story begin-ning and the produc-tion of transcript.