My Favorite Publication
Publication (Recent & Selected)The titles below may not be exactly the same as in the papers. They are used in order to highlight the key ideas/contributions. Click [more] for details of the paper. Click [paper] to download the PDF right away. Click [video] for the project video. Click [software] for the associated software.
Full Publication by Year
Computational Manga
|
Structure-Aware Halftoning
(SIGGRAPH 2008)
Existing halftoning techniques mainly focus on tone reproduction and
annoying pattern avoidance. Visually sensitive fine details are usually
lost during the halftoning. We developed an optimization-based halftoning
method that can preserve the fine texture as well as the tone similarity.
|
|
Self-Animating Images
(SIGGRAPH 2008)
When certain combination of colors and patterns are used in an
image, our human vision system perceives illusion of motion even the
image is static. Based on the limited knowledge of this phenomenon,
we design algorithms to strengthen the illusory effect of motion given
an input image. Such illusion can be utilized in artwork and design.
|
 © Yukito Kishiro / Shueisha |
Manga Colorization
(SIGGRAPH 2006)
The widely popular Japanese manga arts is usually presented in B/W. Its
intensive use of hatching and screening is a distinctive characteristic,
but also imposes many difficulties in its digital colorization. Unlike the
previous colorization methods that rely on intensity-continuity. This
paper presents a colorization technique that relies on the
pattern-continuity as well as the intensity-continuity. Hence it can
naturally colorize the pattern-intensive manga. |
|
Cartoon Deringing
(ACM TOG 2006)
Existing manga (or anime) are normally stored as JPEG (or MPEG). However,
such BDCT encoding is not suitable for images with sharp edges, like
manga/anime. Images are usually contaminated with annoying visual
artifact. This paper presents an analogy-based deringing method to
synthesize artifact-reduced images instead of traditional
postfiltering approaches. Substantially visual and statistical
improvements over existing methods are obtained.
|
Weathering
|
ɣ-ton Tracing
(SIGGRAPH 2005)
ɣ-ton tracing is a visual simulation framework for a wide variety of
weathering effects. In this framework, blemishes (e.g. rust, stain,
patina, etc) are formed by the cumulative propagation of an abstract
weathering-catalysing agent called ɣ-ton. The propagation of
ɣ-ton resembles the photon mapping. Therefore, once you know how to
do photon mapping, you can simulate weathering.
|
|
Geometry-Dependent Surface Imperfections
(EGWR 1997)
The stochastic patterns of real-world weathering phenomena are usually
geometry-dependent. Example geometric factors include surface exposure,
curvature, and external weathering sources (just like the light sources).
By first determining the underlying tendency from
scene geometry, we can simulate a wide variety of weathering phenomena.
|
|
Dust Accumulation
(IEEE CGA 1995)
Although weathering phenomena seems to be stochastic, its underlying
pattern is systematic. Through this specific simulation of dust
accumulation, we demonstrate the weathering pattern is highly related
to geometric factor like surface exposure (or accessibility). |
Sphere Map
|
Isocube
(IEEE TVCG 2007)
The six-face cubemap is widely equipped on GPU. But the cubemap does not
sample the sphere uniformly and each texel does not bear the same weight
(solid angle). Here we propose a new sphere partitioning and sampling
scheme that gives uniform sampling (low discrepancy) and equal
solid-angle properties. Moreover, it fits nicely in the six-face cubemap
hardware and hence it can steal the hardware antialiasing features of
cubemap.
|
Importance Sampling
|
Spherical Q2-tree (EGSR 2005)
To speed up the rendering, we can approximate the distant envrionment
lighting by a finite number of directional lights. The question is how to
sample the spherical environment map. By developing the 360 degree
environment onto a plane with solid angle equality, we then turn the
importance sampling problem into a simple quad-tree partitioning problem
on sphere. |
Relighting and Precomputed Lighting
|
Tileable BTF
(IEEE TVCG 2007)
Dressing an object with BTF normally involves a geometry-dependent
synthesis. The synthesized BTF cannot be reused on another object. Since
the synthesis of high-dimensional BTF is very time-consuming, this hinders
the application of BTF. We proposed the tileable BTF to decouple the BTF
synthesis and the geometry. Hence, once a set of BTF tiles is synthesized,
it can be repeatedly applied on various objects without the lengthy
synthesis. The dressing is instantaneous. |
|
Noise-Proofed SH
(IEEE TVCG 2006)
Storing precomputed lighting information (such as relighting, BTF, and
PRT) with spherical harmonics (SH) can drastically reduce the data size.
However, that is still far away from practical use as each vertex or pixel
has its own SH coefficient vector (of, say, 25 x 25 elements). Further
compression on these SH coefficients may carelessly introduce severe
visual artifacts. We digged up the underlying reason and introduced a way
to noise-proof the SH coefficients, hence practical applications
become feasible. |
|
Dense Photometric Stereo
(IEEE TPAMI)
Given the massive number of images captured for relighting, we can
accurately recover the normal and depth maps using dense photometric
stereo. Recovery is accurate even the lighting direction is not
precisely measured and there are substantial amount of shadow and specular
highlight (which causes much troubles to traditional shape recovery
algorithms). Note that the same set of images is used for depth
recovery and relighting. No range scanner is required. |
|
PCA Compression for Relighting
(IEEE CSVT 2005)
Previous methods in compression relighting data mostly rely on SH which may
smooth out high-frequency details. Here, we propose a block-based PCA
compression scheme. It essentially separates the data into basis images and
relighting coefficients, which can be further effectively compressed using
transform coding and quantization. The nice feature of PCA is its ability in
preserving high-quality shadow and highlights, while maintaining high
compression ratio and fast rendering.
|
|
Compression for Relighting
(IEEE CSVT 2003)
Relighting or precomputed lighting is actually a compression problem of
tons of images. One may straightforwardly compress them with image or
video compression standards such as JPEG, JPEG2000 and MPEG. However, they
do not work well nor fully exploit the data correlation. Unlike the
frame-by-frame playback of video, relighting may randomly access the
pixel values during image synthesis. This paper presents a
divide-and-conquer compression approach which substantially outperforms
existing image and video coding methods. |
|
Real-Time Relighting
(Graphics Programming Method 2003)
Relighting in local illumination (directional and point light sources) is
basically a linear combination of basis functions, either image-wise or
pixel-wise. The grid nature of image-based relighting fits nicely into the
current SIMD-based GPUs and facilitates its real-time rendering.
However, SH imposes difficulties due to its recursive definition. To solve
it, we utilize hardware cubemaps.
|
|
Sampling Bound of Relighting
(CVPR 2001) (IJCV 2002)
How many samples we should take along the lighting dimension in order to
synthesize realistic images? If we know the answer, we can avoid wasteful
oversampling and undesirable undersampling, during the relighting and
precomputed lighting. This is exactly what this paper trying to answer. We
went through an indepth analysis and obtained the theoretical sampling
bound for relighting and precomputed lighting. To verify our bound, we
even build a machine to precisely position the point light source. |
|
The Plenoptic Illumination Function
(IEEE TMM 2002)
The original formulation of the plenoptic function does not include
illumination. Previous image-based virtual reality applications simply
assume that the illumination is fixed. In this paper, we proposed a new
formulation of the plenoptic function, called the plenoptic illumination
function, which explicitly specifies the illumination component.
Techniques based on this new formulation can be extended to support
relighting as well as view interpolation. This paper received the IEEE
Transactions on Multimedia Prize Paper Award 2005. |
|
ABRDF and Image-based Relighting
(EGWR 1997)
The earliest paper on image-based relighting from our group in 1997. We
attempted to include the lighting into the image-based rendering framework.
So that we are able to change the viewpoint as well as the lighting.
Our work is later commonly referred as image-based
relighting. In this paper, we introduced the concept of
apparent BRDF (ABRDF) of pixel, which is the aggregate function
of surface reflectance, shadowing, and other indirect illumination
contribution.
|
View Interpolation
|
Binary-Space-Partitioned Images (IEEE TVCG 2004)
Can we resolve the visibility without
referring to 3D? This paper tries to answer this question. By binary-space
partitioning the input 2D image and projecting the novel viewpoint onto the
view plane of the input image, novel views can be generated. The 3D
visibility can be correctly and completely resolved in 2D. All we need is a
2D BSP visibility sorting. |
|
View Interpolation Without Depth-Buffering
(JGT 1998)
Using epipolar geometry, a visibility-correct drawing order of pixels
can be derived without depth-buffering. The drawback is to do it
pixel-by-pixel. Again, using epipolar geometry, we derive a
visibility-correct drawing order for 2D triangles. The order between
each pair of neighboring triangles is first obtained. Then a graph can be
built and topological sorting is applied on the graph to obtain the complete
drawing order of all triangles in linear time.
|
GPGPU
|
DWT on GPU
(IEEE TMM 2007)
Discrete wavelet transform (DWT) has been widely used in recent years,
e.g. JPEG2000 image coding standard. But DWT is rather computational
intensive when the data size is large. Since most PCs come with a GPU, why
don't we turn it into a powerful DWT engine? As there are many DWTs, we
proposed a method to support various separable DWTs with various boundary
extension schemes. This project started in 2003 and its source
code has been released since 2004. |
|
Evolutionary Computing on GPU
(IEEE IS)
Genetic algorithm is a kind of evolutionary algorithms. Evolutionary
algorithms are effective and robust methods for solving many practical
problems such as feature selection, electrical circuits synthesis, and data
mining. However, they may execute for a long time, because enormous fitness
evaluations must be performed. In this paper, we proposed the first
evolutionary algorithm on GPU, affordable by most people. |
Halftoning
|
Details-Preserving Halftoning
(Graphics Gems V, 1995)
Halftoning is heavily used in printing
industry. Classical methods usually exhibit observable patterns. Here, we
make use of the space-filling curves to avoid such pattern. The
space-filling curve turns a 2D problem into 1D. Hence 1D signal processing
techniques can be easily and efficiently applied to retain the details in
the original grayscale input. |
Visualization
|
Adaptive Skeleton Climbing
(Eurographics 1998)
Generating isosurfaces from volume data
normally results in large amount of triangles. One can use post-processing
mesh decimation technique to reduce the triangle count, but with the
trade-off of long processing time. To address this problem, we proposed an
isosurface extraction algorithm that generates decimated mesh
directly from the volume in a short period of running time.
|
|
Virtual Arthroscopic Surgery
(IEEE TITB 2004)
Minimally invasive surgery requires excellent skill of hand-eye coordination
and instrument manipulation. However, surgery training with cadavers is more
and more difficult. Here, we built a hardware knee arthroscopy surgery
simulator, not just simulate the operational instruments but also the
haptics feedback. |
 |
| Home |
