You are on page 1of 2

Spandan Ghosh 11 Yash Margaje 16 Akshay Kanade 17 Shamika Mhatre 19 Durva Kavathkar 22 Shreyasi Patil 35 Parth Gadekar 36 Suyog

Morankar 40

DEEP LEARNING TECHNIQUES FOR SUPER-


RESOLUTION IN VIDEO GAMES
Dept. of Information Technology,
A. C. Patil College of Engineering,
Kharghar.Mumbai-410210
Source : Deep Learning Techniques for Super-Resolution in Video Games
Abstract: The computational cost of video game graphics is This makes games more accessible whilst improving
increasing, and graphics processing hardware is struggling to performance on high end systems.
keep up. This means that computer scientists must be creative The leader in a consumer solution to this problem is NVIDIA
in new ways to improve the performance of the graphics with the "Deep Learning Super Sampling (DLSS)" rendering
processing hardware. The super resolution of video enables method which claims a 2-3 times performance improvement at
video games to enjoy high quality graphics and offsets a large 4K. Alternatively, there are promising open-source solutions
part of the computational costs. These recent technologies from companies like AMD with their "FidelityFX" technology
enable consumers to improve the performance and enjoyment and Microsoft with their upcoming application programming
of video games, and the has the potential to become the interface (API) called DirectML which they have demonstrated
standard within the game development industry. to upscale videos from 540p to 1080p in real time.

Keywords : deep learning, super resolution, video games, QUESTIONS


neural networks There are questions to be addressed with this emerging
technology which will determine it’s potential to become the
INTRODUCTION industry standard used by the majority of major game
Video game graphics are improving. Complex physics and ray- development studios. What are the percentage increases in
traced light effects, where environmental lighting is rendered in frame rate using deep learning video SR? How accurate is the
real time as individual rays, have become the standard within upscaled image compared to the original? Additional
major game development studios. For users to play games with Performance cost? Hardware required to make use at a
these graphics on their computers, they need dedicated consumer level and how accessible are these solutions for the
hardware called graphics processing units (GPUs). At the same game developers?
time increases in GPU processing power are beginning to They can be answered by looking at the research for deep
plateau as the size of transistors reach the atomic level. learning super resolution by Wang et al
Traditionally performance was improved by shrinking the size
of transistors on a microchip and doubling the count. Now FRAME RATE
computer scientists are devising new methods to improve the The most substantial source of data on frame rate increase using
performance of GPUs using emerging tech such as the use of deep learning video SR is NVIDIA’s official Deep Learning
deep neural networks to enhance video game graphics in real Super-Sampling (DLSS) information page, they display figures
time through a technique called Video Super Resolution (SR). showing the average frames per second (fps) with and without
This has the potential to reduce the GPU performance required DLSS on four different videogames with the frame rate for
to play video games while improving the graphics quality every resolution on each of the tested GPUs. At 1080p the fps
according to Wang et al. increases by a range of 51-63 across the various GPUs, with the
smallest increase coming from the then flagship RTX 2080 Ti
BACKGROUND GPU and an average of 77.8 fps increase from the cheapest
Gaming enthusiasts appreciate the stunning graphics as well as RTX 2060 GPU going from an average of 41.8 fps with DLSS
the high performance of their games, which is a difficult balance off to 68.1 fps with At 4k the frame rate increase is drastically
to maintain. There is a performance cost to playing video games higher with an increase ranging from 159-360 across the GPUs,
at resolutions like 4k, with lots of graphics effects. It can reduce with the RTX 2080 Ti increasing from an average of 25.8 fps
the frame rate to a level at which the game cannot be played in
a pleasant way. Fast-paced action games require a high frame
rate for a smooth gaming experience, especially in competitive
scenarios where fractions of a second can decide between
victory and defeat. Research by Claypool et al show that in a
first-person shooter game, players performed 7 times better at
60 fps compared to reduced frame rate of 3 to 7 fps.
While consumers can increase performance by using lower
quality graphic settings, it is undesirable for both the player and
game developer as their product is not being experienced as
intended. Using Deep Learning Video SR allows consumers
with lower performance GPUs run games at a higher perceived
visual quality and performance than what is usually possible.
From this data we can see that the FPS gains are sizeable, with NVIDIA, "DLSS 2.0 is now available to Unreal Engine 4
percentage increases from games with lower initial average FPS developers through the DLSS Developer program, which will
with DLSS disabled. The results show that Deep Learning SR accelerate deployment to one of the world's most popular game
can be highly effective in improving game performance with engines". Unreal Engine 4 and one of the most popular game
NVIDIA's DLSS, it is able to make the cheaper RTX 2060 GPU engines, which means that these tools are widely used
perform better than their flagship RTX 2080 Ti. internationally. Additionally, AMD's FidelityFX and
Microsoft's DirectML are open-source tools that anyone can
ACCURACY use, although they are still in their infancy compared to
Convolutional neural networks: They are an efficient NVIDIA's solution.
technique for super-resolution deep learning. CNN requires less
pre-processing than traditional SR imaging techniques. it can CONCLUSION
be trained to identify and distinguish between key features in There is the potential for deep learning super resolution
each frame. For SR-Video, CNN can receive the neighbouring techniques to become the standard in the video game industry,
frames as input, it helps the neural network learn the spatial and allowing consumers to achieve substantial increases in
temporal information of the frame, enhancing CNN's ability to performance from their GPUs while maintaining high images.
produce high quality results These emerging technologies are still in their infancy and there
Motion Compensation: By providing motion vectors to CNN, is currently only NVIDIA Deep Learning Super Sampling
the algorithm can compensate for motion between frames and (DLSS) support for a limited number of games. The adjacent
produce high quality SR video images and motion vectors available as input to CNN allow
Deep Learning Super Sampling: NVIDIA's method for SR greater accuracy of the resulting video, with satisfactory results
video takes low resolution alias frame images and their motion based on subjective AQI methods. The performance cost of the
vectors, which contain information about the direction of deep learning algorithms is negligible, with the performance
movement of all pixels in the following frame, and puts them gains exceeding the resulting slight side grade in perceivable
into a CNN, called a convolutional auto encoder, which is graphics quality.
trained on a supercomputer to use these inputs to generate a
high-resolution version of the frame image on a frame-by- FUTURE WORK
frame basis These questions could be further answered with future research
Image Quality Assessment: The results of Deep Learning SR that could be conducted by independently testing NVIDIA's
can be assessed using performance metrics called Image DLSS solution with Microsoft and AMD solutions in a fair
Quality Assessments (IQAs). Objective IQA methods use environment that could control variables - a single-blind study,
computer modelling to compare images. The problem with conducted on a large number of subjects and which could
these methods is that the computer does not perceive as a human provide a fair analysis of any of the video super resolution
does, so the results of the assessment can easily contradict the methods based on observation without bias.
results of the subjective IQA. This makes it difficult to calculate
the performance measurement task effectively, but SR video REFERENCES
A. G. E. Moore, "Progress in digital integrated electronics [Technical literature,
with deep learning can provide effective results based on Copyright 1975 IEEE. Reprinted, with permission.Technical Digest. International
subjective and objective IQA methods. Electron Devices Meeting, IEEE, 1975, pp. 11-13.]," IEEE Solid-State Circuits
Society Newsletter, vol. XI, no. 3, pp. 36-37, 2006.
B. R. Williams, "What's next? [The end of Moore's law]," Comping in Science &
PERFORMANCE COSTS Engineering, vol. XIX, no. 2, pp. 7-13, 2017.
An advantage of SR video over single-frame SR is that adjacent C. Z. Wang, J. Chen and S. C. Hoi, "Deep learning for image super-resolution: A
fields can be used with the current field and motion vectors to survey," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. II,
2020.
reduce the computation required to generate the current field. D. M. Claypool, K. Claypool and F. Damaa, "The effects of frame rate and resolution
This dramatically increases the speed of the algorithm with on users playing first person shooter games," Proceedings of SPIE - The
International Society for Optical Engineering, vol. 6071, 2006.
negligible performance loss. E. NVIDIA DLSS 2.0: A Big Leap In AI Rendering | GeForce News | NVIDIA
F. Radeon™ Software | FidelityFX | AMD
ACCESSIBILITY G. microsoft/DirectML: DirectML is a high-performance, hardware-accelerated
DirectX 12 library for machine learning. DirectML provides GPU acceleration for
Consumers: The only widely available solution for the common machine learning tasks across a broad range of supported hardware and
consumer at present is NVIDIA's DLSS which requires drivers, including all DirectX 12-capable GPUs from vendors such as AMD, Intel,
NVIDIA, and Qualcomm. (github.com)
dedicated specific hardware in the form of one of the RTX
H. Kappeler, S. Yoo, Q. Dai and K. Katsaggelos, "Video super resolution with
series graphics processing units (GPUs). This price is only for convolutional neural networks," IEEE Transactions on Computational Imaging, vol.
the GPU which is only one component of a personal computer II, no. 2, pp. 109-122, 2016.
I. X. Tao, H. Gao, R. Liao, J. Wang and J. Jia, "Detail-revealing deep video super-
(PC), which makes the total price of the entry much higher. This resolution," IEEE International Conference on Computer Vision (ICCV), pp. 4482-
means that the DLSS solution is currently only available to 4490, 2017.
J. Z. Zhang and V. Sze, "FAST: A framework to accelerate super resolution
enthusiastic PC owners with mid to high end PCs. However, as processing on compressed videos," in Proceedings of the IEEE Conference on
this hardware is replaced with new hardware, prices will drop, Computer Vision and Pattern Recognition Workshops, 2017.
making the more accessible to consumers on reduced budgets. K. W. Shi, J. Caballero, F. Huszár, J. Totz, A. P. Aitken, R. Bishop, D. Rueckert and Z.
Wang, "Real-time single image and video super-resolution using an efficient sub-
Developers: For deep learning video SRs to become a standard pixel convolutional neural network," 2016 IEEE Conference on Computer Vision
technique for game development, these techniques must be and Pattern Recognition, pp. 1874-1883, 2016.
L. Christopoulou and S. Xinogalos, "Overview and comparative analysis of game
widely available to game developers worldwide. According to engines for desktop and mobile devices," International Journal of Serious Games,
vol. IV, pp. 21-36, 2017

You might also like