I am a PhD candidate at the computer graphics department of the Max-Planck-Institute for Informatics in Saarbrücken.
My research focuses on interactive Global Illumination algorithms and related areas, in particular Indirect Illumination, Ambient Occlusion, and Shadow Mapping.
For now, I mainly focus on real-time algorithms, running on modern GPUs.
My research advisors are
Elmar Eisemann and
Karol Myszkowski.
03/2012
GPU Pro 3 is out! The book contains our section about Screen-space Bent Cones. The sample implementation has a minor bug in the pre-convolution computation of the environment map, you can download a fixed version from *below*.
10/2011
The
cover of the current
Informatik Spektrum was created with our
Bent Normals and Cones in Screen-space (VMV 2011).
Die Umgebungsverdeckung (Ambient Occlusion) ist eine beliebte Beleuchtungstechnik in der 3D Computergrafik für die photorealistische Darstellung virtueller Szenen.
Es ist eine vereinfachende Näherung für Umgebungslicht (environmental lighting), wobei der Eindruck von globaler Beleuchtung entsteht.
Die Technik beschleunigt dabei aufwändige Lichtberechnungen durch eine Trennung von Sichtbarkeit und Lichtreflektion.
Hierbei wird das Licht aus allen Richtungen aggregiert und mit der mittleren Sichtbarkeit eines Oberflächenpunkts, ebenfalls aus allen Richtungen, multipliziert.
Dies führt jedoch zu einer Reduzierung des Photorealismus, da die Beleuchtung richtungsabhängig ist.
Durch Verdeckung verursachte Schatten sind grau und werden als eine Änderung von Materialeigenschaften, nicht jedoch des Lichts wahrgenommen.
Im Bereich der Filmindustrie wurden geneigte Oberflächennormalen (Bent Normals) als ein Lösungansatz eingesetzt.
In der Arbeit wird beschrieben, wie Bent Normals und eine neue Erweiterung, Bent Cones, im Bildraum berechnet werden können.
Anstatt nur die mittlere Sichtbarkeit zu bestimmen, wird gleichzeitig die gemittelte sichtbare Richtung berechnet.
Zusätzlich lässt sich die Verteilung der sichtbaren Richtungen direkt aus der nicht-normalisierten, mittleren Richtungen ableiten.
Diese beiden zusätzlichen Daten ermöglichen die Bestimmung eines Sichtkegels.
Während der Lichtberechnung wird nun nur jenes Licht aggregiert, welches innerhalb des Sichtkegels liegt, Licht außerhalb wird ignoriert.
Dies ermöglicht farbige Schatten, wenn Licht aus unterschiedlichen Richtungen kommt.
Ebenso folgt die Form der Schatten nun der Richtungsabhängigkeit der Beleuchtung.
Ambient Occlusion wird weiterhin zur Abschätzung der mittleren Sichtbarkeit verwendet, was auch dann eine plausible Abdunklung sicherstellt, wenn der Kegel die tatsächliche Sichbarkeit nur schlecht annähert.
In diesem Fall wird Licht wieder aus allen Richtungen aggregiert, das bedeutet ein Rückfall auf die Beleuchtung nur mit Ambient Occlusion.
Die ähnliche Methode „Screen-space Directional Occlusion“ erzielt vergleichbare Ergebnisse, bricht jedoch das Prinzip der Trennung von Sichtbarkeit und Lichtreflektion, weshalb es nur halb so schnell ist.
Unsere Technik zur Bestimmung der Bent Normal und des Bent Cones verbindet die Berechnungsgeschwindigkeit und Einfachheit bestehender Ambient Occlusion Verfahren mit einer realitätsnäheren Lichtberechnung.
Dabei kann sie leicht in bereits vorhandene Bilderzeungssysteme integriert werden.
Da die Berechnung in wenigen Millisekunden auf moderner Grafikhardware durchgeführt wird, sowie Möglichkeiten zur Abwägung zwischen Qualität und Berechnungsgeschwindigkeit vorhanden sind, eignet sich die Technik für interaktive Anwendungen, insbesondere Computerspiele.
Weitere Details können dem Paper „Bent Normals and Cones in Screen-space: Proceedings Vision, Modeling and Visualization Workshop 2011 (Berlin, 04.10.2011 - 06.10.2011)" entnommen werden.
10/2011
I presented our paper about Bent Normals and Cones in Screen-space at VMV in Berlin. There were some quite interesting papers.
I got some nice feedback about our paper. People were surprised about the nice results, although the method is rather simple.
08/2011
We wrote an article
Screen-space Bent Cones: A Practical Approach, which will be part of the upcoming
GPU Pro 3 book.
As we think our technique is easy to implement, while improving on the staet-of-the-art SSAO, the topic should be of value for the target audience of the GPU Pro book series: game developers.
We explain the technique more from a developer point of view and show possibilities for tweaking like trading quality for performance. This is topped off by code samples of our implementation.
07/2011
Our paper about
Bent Normals and Cones in Screen-space has been accepted at
VMV 2011.
Basically, SSBC (screen-space bent cones) extend SSAO (screen-space ambient occlusion) by considering the directionality of light, while keeping the computation as fast and simple as SSAO.
This enables colored shadows and the shadows shape follows the directionality of illumination.
In direct comparison to
Screen-space Directional Occlusion (Approximating Dynamic Global Illumination in Image Space, Ritschel et al.) the new technique requires less samples, achieving almost double the performance.
2011
Screen-space Bent Cones: A Practical Approach
O. Klehm, T. Ritschel, E. Eisemann, H.-P. Seidel
GPU Pro 3, editor: Wolfgang Engel, A K Peters
Demo & Code (look at canvas.cpp for hard-coded parameters you want to change)
updated: Friday, 02 March 2012 - 15:38; size: 9.7M
Ambient occlusion (AO) is a popular technique to visually improve both real-time as well as offline rendering.
It decouples occlusion and shading providing a gain in efficiency.
This results in an average occlusion that modulates the surface shading.
However, this also reduces realism due to the lack of directional information..
Bent normals were proposed as an amelioration that addresses this issue for offline rendering.
Here, we describe how to compute bent normals as a cheap by-product of screen-space ambient occlusion (SSAO).
Bent cones extend bent normals to further improve realism.
These extensions combine the speed and simplicity of AO with physically more plausible lighting.
@inbook{SSBC_GP3_Klehm11,
title={Screen-space Bent Cones: A Practical Approach},
author={Oliver Klehm and Tobias Ritschel and Elmar Eisemann and Hans-Peter Seidel},
editor={Wolfgang Engel},
booktitle={GPU Pro 3},
series={GPU Pro}
year={2012},
pages = {191--207},
publisher={CRC Press}
}
Bent Normals and Cones in Screen-space
O. Klehm, T. Ritschel, E. Eisemann, H.-P. Seidel
16th International Workshop on Vision, Modeling and Visualization
4-6 / October / 2011, Berlin
updated: 07/10/2011, now with correct performance comparison
Ambient occlusion (AO) is a popular technique for real-time as well as offline rendering.
One of its benefits is a gain in efficiency due to the fact that occlusion and shading are decoupled which results in an average occlusion that modulates the surface shading.
Its main drawback is a loss of realism due to the lack of directional occlusion and lighting.
As a solution, the use of bent normals was proposed for offline rendering.
This work describes how to compute bent normals and bent cones in combination with screen-space ambient occlusion.
These extensions combine the speed and simplicity of AO with physically more plausible lighting.
@inproceedings{SsbcVMVKlehm2011,
AUTHOR = {Klehm, Oliver and Ritschel, Tobias and Eisemann, Elmar and Seidel, Hans-Peter},
EDITOR = {Eisert, Peter and Hornegger, Joachim and Polthier, Konrad},
TITLE = {Bent Normals and Cones in Screen-space},
BOOKTITLE = {16th International Workshop on Vision, Modeling and Visualization},
PUBLISHER = {Eurographics Association},
YEAR = {2011},
PAGES = {177--182},
ADDRESS = {Berlin, Germany},
MONTH = {October},
ISBN = {978-3-905673-85-2},
}
2010
Interactive Massive Lighting for Virtual 3D City Models
O. Klehm
Master's Thesis
Supervisors: Prof. Dr. Jürgen Döllner and Dr. Haik Lorenz
Hasso-Plattner-Institute, University of Potsdam, Germany
Photorealistic rendering presents a main approach for depicting virtual models and worlds.
In photorealistic rendering, illumination is a key functionality to achieve appealing and plausible visual results.
Due to the diversity of effects and global nature of light, interactive applications require a simplification of illumination (commonly to local illumination only considering direct illumination).
But even with this limitation the number of light sources remains restricted for arbitrary scenes.
Forward rendering is the standard way of applying lighting, the operation of calculating illumination, which is performed for each object.
As such, the complexity of lighting depends on scene complexity. However, virtual scenes constantly increase in complexity by pursuing more realism.
This development also increases lighting costs and thus limits a similar growth in the number of light sources.
One solution is separating lighting from scene complexity.
This thesis describes two approaches to decouple lighting from geometry processing.
Both techniques, deferred shading and deferred lighting, render a view dependent scene description into an image buffer (so called G-Buffer).
Lighting is performed in a \textit{deferred} stage in image space with geometry data read from the image buffer.
We limit the light distribution of a light source to a volume and thus perform lighting for important pixels only.
Users can trade-off quality and performance by defining a threshold to extend or shrink the light volume.
As an important application of the algorithms, we implement an indirect illumination approach by [Dachsbacher and Stamminger, 2006], which builds on the idea of instant radiosity [Keller, 1997].
It turns indirect illumination into direct illumination by placing virtual point lights (VPLs) at directly lit surfaces.
Respective properties of the VPLs are determined by rendering a reflective shadow map [Dachsbacher and Stamminger, 2005] from the light's point of view.
Depending on the scene, thousands of VPLs are necessary to appealingly add indirect illumination.
The technique integrates well with the deferred shading/lighting approaches, taking advantage of their optimizations.
We further implement shadow mapping [Williams, 1978] as proof of concept and a post processing pipeline including gamma correction and high dynamic range rendering.
All these algorithms are integrated into a generic lighting library, which provides interfaces for the OpenSceneGraph rendering framework and plain OpenGL contexts.
Deferred shading and deferred lighting are scene and view dependent but allow for multiple magnitudes more light sources than possible with forward rendering.
Thus, we are able to render night scenes of a 3D virtual city model with 20,000 light sources without visual artifacts.
As an extreme case, we demonstrate a sunset scenario with 160,000 VPLs causing smooth indirect illumination, rendered at interactive frame rates.
@mastersthesis{Klehm10,
title={Interactive Massive Lighting for Virtual 3D City Models},
author={Oliver Klehm},
school = {Hasso-Plattner-Institute, University of Potsdam},
year={2010}
}
