Methods for user guided compression algorithms
Jostein Bratlie
R&D Group for Mathematical Modelling,
Numerical Simulation & Computer Visualization,
Faculty of Technology, Narvik University College, Narvik, Norway
Abstract
In computer games of today and specially massively multiplayer online
role playing games virtual world and the characters therein are becoming
more and more realistic, and therefore demands a larger and larger
amount of data to be described in detail. As a part of the ’DreamWorld’
research project, Funcom and Narvik University College have entered
into a co-operation to find new ways to handle some of the issues linked
to this.
This paper gives and introduction to a PhD project which aim is
to develop new methods for user guided compression algorithms. It
also aims to end up with a prototype for a user guided animationcompression-tool for seamless integration into an already existing
production pipeline.
1
Introduction
Virtual worlds are expected to be more and more realistic. With increased artistic
freedom and demand for realism the amount of data describing these worlds
increase exponentially for each released title. Especially with respect to MMORPGs
(massively multiplayer online role playing games). As a title is released, the
expectations for the level of realism in the next title will be heightened. With
increased realism the first thing that comes to mind is visual effects, increased
resolution of models and more stunning scenario. In addition to the sheer volume
of visual effects are the quality of story-telling, dialogs, voice work and acting also
increasing in the same way. This again demands a richer animation base for ingame, non cinematic characters. All these different parts yields larger amount of
data, which has to be stored, processed and distributed. Todays games cinematics
is also being simulated with in-game techniques, which again adds to the size of
animation data. In many cases the combined size of in-game content for a computer
game, again specially for an MMORPG, can reach a respectable size of up to 50100GB. Distributing content of this size by electronic means as digital downloads
may be a bottleneck, demanding download times that scare off potential players.
In addition there is the ever increasing amount of client/server communication, like
This paper was presented at the NIK-2011 conference; see http://www.nik.no/.
more complex movement which generates more positions and directional data that
has to be synchronized as realism increases.
So, there is a certain demand for production tools and software algorithms that
give good compression, while still being easy to use and integrate into existing tool
chains so that the artists will still be in control of the output.
As a part of the ’DreamWorld’ research project (thanks to NFR (Norges
Forskningsråd) and its ’VERDIKT’ program), Funcom and Narvik University
College have entered into a co-operation which give me the possibility to work with
some of these issues. The main focus for my dissertation will be on user guided
compression of animation data.
Possible applications is therefore user-managed methods for compression of
different types of massive data streams. This includes, but is not limited to, signal
and image data as well as geometric data and data of different structures.
2
Transforms and indexing
For these kinds of tools to be interactive it is necessary to optimize this for highly
parallel computation on multi core or GPU (Graphics Processing Unit) architecture.
Some of this work has already been conducted by the research group at Narvik
University College.
Since 2004 Narvik University College has been conducting research on different
topics within this area. Such as wavelet shrinkage: curve and surface fitting, [8],
and strategies, [3]. Visualization and isometric conversion between dimension and
resolution, [4, 5, 7]. As well as computing n-variate orthogonal discrete wavelet
transforms on graphics processing units, [6].
Discrete wavelet transforms, operations on the transformed data and packing
algorithms are usually found at the core of most compression algorithms today.
In an upcoming publication our latest efforts investigating prerequisites for these
algorithms have brought us some interesting new ways of looking at data structures.
An algorithm for indexing blocks of tensor-product wavelet bases has been proposed
in [6], and in [2] the algorithm was extended to directly index the individual bases
inside each of the basis-blocks. In theory this allows for computation of an ndimensional discrete wavelet transform of an n-dimensional data set to be performed
in one dimension. The case of I = 1d 7→ nd and the inverse I −1 7→ 1d. Then in
our latest efforts an extended algorithm building the aforementioned algorithm once
more were built for the general case, so one can map basis functions between any
two dimensions, Id1 7→d2 .
Using a one dimensional DWT (Discrete Wavelet Transform) to compute a higher
dimensional DWT does not only make the implementation easier by giving us a
unified framework for such computations, but some expectations about using them
as operators on data sets that otherwise would not have been seen as easily streamable, will fall.
3
Data approximation and compression
This is the start of a three year long PhD project which will be finalized in the
beginning of 2014.
An Narvik University College in-house open source geometry and graphics
library, GMlib [1], is planned to be extensively utilized and extended for prototyping
during this project. The library has been developed as a collection of geometric
modeling, graphics and computer science tools and algorithms since 1994, and are
now available as an open source library on the Episteme [1] websites of Narvik
University College.
Some tasks is also considered scheduled in the near future. GPU implementation
of a multivariate indexation algorithm is already at a prototype stage with code
for generating example illustration utilizing OpenCL. This will most likely also be
extended under different programming paradigms, as CUDA (NVidia) and Direct
Compute (Microsoft). Prototype implementations for other compression test cases
as JPEG2000 and H.264 are also planned, and will most likely generate considerable
work. A first Funcom pipeline-ready prototype of a working tool based on ”Methods
for user guided compression algorithms” is not expected until late 2012. Pre and
post processing and approximation of data on sinal level using for example Exporational B-splines is somewhat linked to the Funcom pipeline-ready prototype and
will be investigated in parallel.
Figure 1: In the top left the traditional ”Lena” benchmark can be seen, and below
it at the left side a one dimensional representation of the same image. This one
dimensional set have been put through a three level one dimensional forward DWT,
and at the bottom right you can see the one dimensional representation of the result.
When this result is indexed as in [6], we can put it together again as seen in the top
right of the picture, which coincides with the output of a traditional two dimensional
forward DWT.
References
[1] J. Bratlie. Episteme, home of gmlib http://episteme.hin.no, July 2011.
[2] L. Dechevsky, J. Bratlie, and J. Gundersen.
Index mapping between
tensor-product wavelet bases of different number of variables, and computing
multivariate orthogonal discrete wavelet transforms on graphics processing units.
In Large-Scale Scientific Computing, Lecture Notes in Computer Science, 2011.
LSSC 2011, Sozopol, Bulgaria. Avaiting review.
[3] L. T. Dechevsky, N. Grip, and J. Gundersen. A new generation of wavelet
shrinkage: adaptive strategies based on composition of lorentz-type threshold
and besov-type non-threshold shrinkage. In F. Truchetet and O. Laligant,
editors, Wavelet applications in industrial processing V, Optics East 2007, 2007.
Volume 6763 of Proceedings of SPIE. Bellingham, Washington 2007. Paper
676304.
[4] L. T. Dechevsky and J. Gundersen. From dynamical visualization of large 3d
and 4d geometrical data sets of isometric conversion between dimension and
resolution. In Large-Scale Scientific Computing, volume 3 of Preprint, page 21.
Narvik University College, 2004. ISSN 1504-4653.
[5] L. T. Dechevsky and J. Gundersen. Isometric conversion between dimension and
resolution. In M. D. ælen, K. M. rken, and L. Schumaker, editors, Mathematical
methods for Curves and Surfaces, pages 103–114, 2005.
[6] L. T. Dechevsky, J. Gundersen, and B. Bang. Computing n-variate orthogonal
discrete wavelet transforms on graphics processing units.
In I. Lirkov,
S. Margenov, and J. Wasniewski, editors, Large-Scale Scientific Computing,
volume 5910 of Lecture Notes in Computer Science, pages 730–737. Springer
Berlin / Heidelberg, 2010. 10.1007/978-3-642-12535-5 87.
[7] L. T. Dechevsky, J. Gundersen, and A. R. Kristoffersen. Wavelet-based isometric
conversion between dimension and resolution and some of its applications.
In F. Truchetet and O. Laligant, editors, Wavelet applications in industrial
processing V, Optics East 2007, 2007. Volume 6763 of Proceedings of SPIE.
Bellingham, Washington 2007. Paper 67630Q.
[8] T. Moguchaya, J. Gundersen, N. Grip, L. Dechevsky, B. Bang, A. Lakså,
E. Quak, and B. Tong. Curve and surface fitting by wavelet shrinkage using
gm-waves. In M. Dælen, K. Mørken, and L. Schumaker, editors, Mathematical
methods for Curves and Surfaces, pages 263–274, 2005.