1.
ComplexCTTP: Complexity Class Based Transcoding Time
Prediction for Video Sequences Using Artificial Neural Network
Anatoliy Zabrovskiy, Prateek Agrawal, Roland Mathá, Christian Timmerer, Radu Prodan
The Sixth IEEE International Conference
on
Multimedia Big Data
September 24-26, 2020 New Delhi.

2.
Motivation
Current situation:
● Many video codecs (AVC, HEVC, VP9, AV1 and etc.)
● Transcoding time depends on many technical aspects (content complexity, transcoding
parameters, processing units)
● Transcoding of video segments is a parallel process running on a high-performance infrastructure
such as the cloud
❏ Transcoding services and platforms work without any prediction of the transcoding time
❏ Transcoding time prediction can significantly improve the overall transcoding time
2

3.
ComplexCTTP method. Goal
Goal:
- Accurate transcoding time prediction for video sequences
The approach is based on two phases:
- Data generation
- Transcoding time prediction using ANN
3

4.
Dataset
2 codecs
x
19 bitrates
x
240 segments
x
9 encoding presets
=
82080
transcodings
(580 hours)
4
Original video file
characteristics
Average spatial information (SI) and temporal
information (TI) for original video sequences
Transcoding with
FFmpeg
Raw transcoding
dataset
Video codecs: x264 and x265
Bitrates:
Segments: 160 (2 sec), 80 (4sec)
Encoding presets: ultrafast,
superfast, veryfast, faster, fast,
medium, slow, slower, veryslow
Training/testing
datasets
We performed the transcoding on a Intel
Xeon Gold 6148 2.4 GHz processor
Contains the maximum and
minimum transcoding time for
all possible combinations of
- codec type,
- complexity class,
- encoding bitrate,
- encoding preset,
- segment duration,
- fps

5.
Segment 2
Segment 1
Segment 3
Segment n
Segment complexity classification and ANN
Original video
segments
Segment 2
Segment 1
Segment 3
Segment n
Segments with low
resolution (144p)
and bitrate
Calculating
SI and TI
(per segment)
Calculating
Complexity class
(per segment)
5
Encoding segments to low
bitrate and resolution
The correlation coefficient between
encoded video segments with 144p
resolution and the original video segments
with 2160p resolution presents positively
strong (0.98 for TI) and highly correlated
(0.65 for SI) relationship.

6.
Results and analysis. Actual transcoding time
Average actual transcoding time for all segments
belonging to a particular complexity class
- x265 requires more more computing
resources than x264
- Segment transcoding time depends on
the complexity of the content
- The transcoding time increases with the
complexity class of the content
- Сomplexity class significantly describes
the complexity of the video segments in
terms of the time required for transcoding.
6

7.
Results and analysis. ANNs
Based on the results:
- ANN with complexity class (1) input parameter predicts transcoding time better compared to the (2)
ANN model without any complexity class, TI and SI input parameters or ANN model with TI and SI (3)
of segments transcoded with a low resolution and bitrate.
- The ANN model with TI and SI input parameters of original video segments (4) has slightly better MAE for both
x264 and x265 codec compared to our ANN model with complexity class. Unfortunately the calculation of TI
and SI metric for the original video segments with high bitrate and resolution takes more time.
7

8.
Results and analysis. ComplexCTTP method vs Tewodros et al.
OVCTT dataset:
- Less transcodings
- Lower maximum values and standard deviations
of transcoding times
- Includes transcodings for MPEG-4 Part 2, VP8
and H.263 codec.
8
OVCTT - Online Video Characteristics and Transcoding
Time Dataset
ComplexCTTP dataset outperforms OVCTT dataset for almost all characteristics!
Transcoding time characteristics of both datasets Transcoding parameters characteristics of both datasets

9.
Results and analysis
9
The Tewodros et al. use bitrate, framerate, resolution, codec, number and size of I, P, B frames as
input parameters for their ANN model.
The average time (in sec) required to calculate ANN input
parameters for one beauty video 2s segment using both
the methods.
Percentage decrease of time (PDT) for 2 sec.
segments for Beauty video sequence is about 70%
Percentage decrease of time for all ten
video sequences with 4s segments.
PDT values range from 53% to 80%

10.
Results and analysis
10
With our ComplexCTTP method, we were able to
minimize MAE to 1.37 for AVC/x264 which is an
improvement of approximately 22% as compared
to the Tewodros et al. method (MAE 1.76).
The result shows that ComplexCTTP performs
better in terms of prediction accuracy.
Coefficients of determination for both the methods

11.
Conclusions
● We proposed video complexity classification, with respect to the video segment’s
spatial and temporal information
11
● We introduced a fast approach to measure SI and TI
● The developed ANN model is able to predict the video transcoding time with low
mean absolute error.

12.
Future work
● Experiments on new emerging codecs
● Using the predicted transcoding time for the actual scheduling of video
transcoding tasks
● Intelligently selecting and analyzing the content complexity of a few segments
of a video to make prediction about the transcoding time of the entire video
12

13.
Thank you!
13
Anatoliy Zabrovskiy
anatoliy.zabrovskiy@aau.at