A New ContentBased Hybrid Video Transcoding Method

1
A New Content-Based Hybrid Video Transcoding
Method

• YongQing Liang
• YapPeng Tan
• Presented by Robert Hung

2

• Introduction
• Video Content Descriptors
• Proposed Selection Method
• Experimental Results
• Summary

3
Introduction

• Aim Bitrate reduction of a compressed video.
  Construct a system consists of three common
  reduction techniques in transform coding
• Requantization (RQ)
• Spatial resolution downsampling (SD)
• Temporal resolution downsampling (TD)
• Problem Integrates the three techniques and
  selected accordingly
• Strategy on the selection of the techniques
• (No mention on why needs three techniques, Not
  addressing the issue on the use of two or three
  techniques together at the same frame, a hole
  needed to be filled, maybe, indeed, no need to
  consider, but why?)

4
Introduction

• Proposed Solution
• Define two descriptors of the video contents as
  the input parameter of the selection method.
• a) Motion activity descriptor (MA)
• b) Spatial activity descriptor (SA)
• The input parameters to the selection method
  including
• two video content descriptors, (MA, SA)
• target bitrate (TB)
• the original frame rate (FR)
• New Selection Method based on some heuristic rules

5
Introduction

• The system is illustrated below

TB
MA
Processed video
Decoded video
SELECTOR
SA
FR
TD
RQ
SD
6
Video Content Descriptors
Motion Activity Descriptor - MA average
magnitude of the motion vector of a frame -
intracoded block has the predefined maximum
motion vector (Not a motion compensation
block) - notcoded block has zero motion vector
(Block has the same pattern as the previous one
at the same location) -AMA average motion
activity over several consecutive frames -200
p-frames statistics shown in the next slide, the
correlation coefficient is 0.92. (only one
set of data)
7
MA against bits of the frame

• MA against the bits of the frame

8
Video Content Descriptors

• Spatial Activity Descriptor
• -SA Mean Quantization of the frame
• - The rationale is that a video frame contains
  a lot of spatial details, more bits are required
  to code the frame with fixed quantization scales.
  If the bitrate is fixed, larger quantization
  scales will be used to code the frame.
• (The measure is used as reference for the next
  frame)

9
Proposed selection Method

• Two main categories of frame
• Low frame rate
• High frame rate
• Some reasoning behind the selection method

10
Proposed selection Method
Reasoning
High spatial activity
High motion activity In the consecutive frame
Rule 1. If RQ can achieve the requirement, do
RQ Rule 2. If Low motion activity, do RQ Rule
3. If motion activity is high, SD is applied.
Otherwise RQ since previously used SD. Rule 4.
If Both spatial and motion activity are high,
SD is applied
Low Frame rate
is the average motion activity of the several
consecutive frames
is the average quanatization scale of the
previous frame
11
Proposed selection Method
the original quantization scale of the current
frame
Target bit rate
High Frame Rate
The actual bitrate of the previous frame
12
Proposed selection Method
Rule 1 Previous selection is on RQ
gt
gt TD selected
is small
Reasons low motion activity spatial detail
can be retained
gt
30
gt SD selected
gt
is high
Reasons SD is selected instead of TD because of
the high motion activity ,
13
Proposed selection Method
Rule 2 Previous selection is on SD
gt
gt RQ selected
is small
Reasons low motion activity spatial detail
can be retained
14
Proposed selection Method
Rule 3 Previous selection is on TD
lt
gt RQ selected
Reasons A little adjustment can achieve the
target bitrate
gt
30
gt SD selected
is high
gt
Reasons SD is selected instead of TD because of
the high motion activity
15
Experimental Results
HVT
Transcode the video of volley ball from
636Kbits/s to 140Kbits/s, HVT is 1dB higher in
average.
16
Experimental Results
Very low PSNR for Requantization
Transcode the video of skating from 112Kbits/s
to 50Kbits/s, HVT is 0.5 and 6.0 dB higher in
average than RQ and SD respectively .
17
Summary

• A method on selecting the bitrate reduction
  techniques
• Simple rules
• Simple measurements
• Good result
• Motion activity and Spatial activity descriptor
  are defined to characterize the video contents

18
Comment

• Rules should be expressed more precisely.
• (If the authors could lay down the Pro and Con
  of each techniques, and derive the conditions for
  employing them.)
• Good results, but the implementation is difficult
  to follow, no mention of which implementation of
  RQ,SD and TD.
• No mention on how it is switching from one
  techniques to another. E.g. how to handle the
  change in resolution between frames.
• It is the current state of art in hybrid video
  transcoder