Exploiting Temporal Coherence

in Ray Casted Walkthroughs

Vlastimil Havran, MPI Informatik
Jir Bittner, Vienna University of Technology
Hans-Peter Seidel, MPI Informatik

Abstract Z-buffer is a powerful algorithm for hidden surface re-

moval, but it can be very inefficient for large scenes with a
We present a technique that aims at exploiting temporal high depth-complexity. In that case the algorithm needs to
coherence of ray casted walkthroughs. Our goal is to reuse rasterize many scene objects even if they are invisible and
ray/object intersections computed in the last frame of the the algorithm performs significant overdraw. The over-
walkthrough for acceleration of ray casting in the current draw problem is commonly addressed by visibility culling
frame. In particular we aim at eliminating the ray traver- methods [14] that aim to detect invisible parts of the scene
sal and computing only a single ray/object intersection per and exclude them from rendering. The goal of these meth-
pixel. If our technique does not succeed in determining ods is to achieve output-sensitivity, i.e. a running time pro-
visibility, it falls back to the classical ray traversal. Vis- portional to the number of visible objects.
ible point samples from the last frame are reprojected to
the current frame. To identify whether these samples can In contrast to the z-buffer algorithm, an efficient im-
be reused we apply splatting and epipolar geometry con- plementation of ray casting performs inherent visibility
straints. We discuss two additional techniques that handle culling. Visibility at the pixel is resolved as soon as the
correct appearance of small objects. We conducted a se- closest ray/object intersection is found. Thus for large
ries of tests on walkthroughs of building interiors. Our densely occluded scenes a software implementation of ray
method succeeded in determining visibility of more than casting can outperform hardware accelerated z-buffered
78% of pixels. For these pixels only a single ray/object rendering even on the latest graphics hardware [29]. It
intersection is executed. The frame rate is increased by has even been forecast that ray casting is likely to sur-
up to 47%. Finally, we argue that the achieved speedup is pass z-buffer in the future for rendering highly complex
relatively significant by comparing the performance of our scenes [28, 29]. Another advantage of ray casting over
algorithm to the ideal ray shooting algorithm. z-buffer is its versatility with respect to the type of prim-
itives it handles. Ray casting can be applied easily for
Keywords: hidden surface removal, ray casting, ray polygons, parametric surfaces such as NURBS, implicit
shooting, reprojection, walkthroughs. surfaces, procedural objects, or CSG models.
The common drawback of ray casting compared to most
1 Introduction other hidden surface removal algorithms is that it does not
exploit coherence of nearby rays. In this paper we ex-
Ray casting was introduced by Appel [2] to solve the prob- ploit temporal coherence between rays shot in subsequent
lem of hidden surface removal already in the late 1960s. frames of walkthrough of a static scene. In our model we
Ray casting solves visibility by shooting rays through each assume that the ray/object intersection and shading should
pixel in the synthesized image and computing the closest be computed for each pixel and each viewing position. The
ray/object intersection. This elementary visibility query goal of the proposed technique is to reduce the overhead
is also called ray shooting [3]. Ray shooting has been of ray casting to computing a single ray/object intersection
thoroughly researched for more than three decades and it per pixel. We also discuss an extension of the method for
has been applied to solve various visibility problems [10]. ray tracing, namely the handling of the point light sources.
Currently the most common application of ray shooting
We use a reprojection of visible points from the pre-
are global illumination algorithms, such as ray tracing or
vious frame to the new frame to eliminate ray traversal.
path tracing [3]. The original application of ray shoot-
The reprojected points are used to determine the object
ing, the hidden surface removal, is nowadays typically re-
that intersect the corresponding ray in the current frame.
solved by the z-buffer algorithm [6] that is commonly im-
We identify problems that occur due to reprojection and
plemented in graphics hardware.
propose techniques that handle them. If the reprojection
{havran, seidel}@mpi-sb.mpg.de does not succeed in determining visibility, we invoke the
bittner@cg.tuwien.ac.at ordinary ray traversal algorithm.
2 Previous Work
Ray casting was introduced by Appel [2] to address the
hidden surface removal problem. Since then many tech-
niques have been proposed to accelerate the elementary
query that is often called ray shooting [3]. A naive imple-
mentation of ray shooting that tests all scene objects for
an intersection with the given ray exhibits (n) time com- Figure 1: Dilatation of the object by 3 3 kernel.
plexity. Szirmay-Kalos [27] has shown that the worst case
lower bound of the ray shooting is (log n). Havran [13]
shows experimentally that ray shooting based on kd-trees P OSi (Points in Object Space). Each entry of P OSi is a
exhibits such a logarithmic behavior in average case. triple (P, O, t) addressable by pixel coordinates (x, y). P
Ray shooting acceleration techniques typically use a represents coordinates of a point in 3D space, O is a refer-
spatial data structure to accelerate the ray traversal and to ence to the intersected object, and t is the signed distance
reduce the number of ray object intersections [3]. Tem- along the ray to the object from center of projection Ci .
poral coherence between frames of an animation can be Another 2D array AIPi (Auxiliary Image Plane) is used
used to further improve the ray tracing performance. Su- to maintain reprojected points for the i-th frame. Each en-
darsky [26] surveyed rendering techniques that exploit try of AIPi contains a triple (O, t, cnt). O is the reference
temporal coherence, more recent paper is by Demers et to an object that is reprojected from the object space to the
al. [9]. Temporal coherence to accelerate rendering has pixel, t is the smallest signed distance from the camera
been studied in the context of volume rendering [12, 31, center to the object reprojected to the pixel, and cnt is the
15], visibility culling [8, 11, 5], and global illumina- number of points reprojected to the pixel (x, y) of AIPi .
tion [30, 24, 25, 17, 21].
The reprojection techniques are also commonly used in 3.2 Reprojection
the context of point-based rendering [20, 22]. Our method
is closely related to work of Badt [4] who was the first to The reprojection constructs the AIPi array by reproject-
use reprojection for ray tracing acceleration. Our method ing points of the P OSi1 array to the current viewing
is also similar to the reprojection technique of Adelson plane V Pi . For each entry of P OSi1 we reproject the
and Hodges [1] which approximates visibility and shad- corresponding 3D point by computing coordinates of cor-
ing by reprojecting already shaded points and rechecking responding pixel (x, y) in AIPi and the signed distance t
the visibility of primary rays explicitly. In contrast to the from current camera position Ci.
approaches mentioned above we aim at computing correct For each projected point we overestimate the influence
visibility and perform view-dependent shading. We elimi- of its projection by splatting using a 3 3 kernel: We
nate artifacts introduced by the reprojection and introduce examine the 3 3 neighborhood of a pixel (x, y). If no
heuristics that handle the undersampling problem. object has been reprojected to a pixel (ix, iy) (ix = x
1, x, x + 1, iy = y 1, y, y + 1), we store the attributes
of the reprojected point to AIPi at (ix, iy). If the same
3 The New Algorithm object has already been reprojected to the AIPi at (ix, iy),
we increment the counter AIP (ix, iy).cnt. In this way
The proposed technique can be subdivided into the repro- we count how many points from 3D space from the same
jection phase and the rendering phase. In the first frame object were reprojected to the pixel (ix, iy). If a point
we only invoke the rendering phase by using an ordinary from another object has already been reprojected to pixel
ray shooting algorithm (ORSA). For computation of sub- (ix, iy) and if its signed distance is bigger than t for the
sequent frames we first apply the reprojection phase fol- currently reprojected point, we overwrite the attributes at
lowed by the rendering phase. We first discuss elementary AIPi (ix, iy) by the attributes of the reprojected point.
data structures used in our algorithm followed by the dis- In the case of reprojecting a single object the splatting
cussion of the reprojection and the rendering phases. leads to a dilatation of the object illustrated in Fig. 1.

3.1 Data Structures 3.3 Rendering

For the further discussion we use a common perspective In the rendering phase we use the results of the reprojec-
camera model. We assume that for the i-th frame of a tion to decide if it is sufficient to test the ray for intersec-
walkthrough we generate an image on a viewing plane tion with a single object associated with the corresponding
V Pi with respect to the center of projection Ci. entry of AIPi , or if we have to apply the ORSA that con-
The algorithm uses two auxiliary arrays of the same di- siders all scene objects.
mension as the image. All ray/object intersection points For each pixel, the AIPi contains an information about
computed in the i-th frame are stored in an array denoted the number of points from the closest object that are repro-
jected to the 3 3 neighborhood of the pixel. We compare Algorithm 2 Rendering Phase.
this number with a constant (mcnt) to decide if ORSA Initialize P OS.
can be avoided. If the number of reprojected points is for each pixel (x, y) on the image plane do
greater than mcnt we compute an intersection of the given Initialize Ray(x, y)
ray and the object associated with the corresponding entry if AIP (x, y).cnt mcnt then
of AIPi . Otherwise we invoke the ORSA. We found out if (Ray(x, y) intersects AIP (x, y).O) then
experimentally that setting mcnt = 5 (as a majority from {Intersection is found.}
9) the reprojection did not introduce any error at this step. end if
The following two pseudocodes outline the reprojection end if
and the rendering phases of the proposed method. if Intersection has not been found by reprojection
then
Algorithm 1 Reprojection Phase. Compute result for Ray(x, y) by ORSA.
Initialize all AIP items (O=0,t=,cnt=0). end if
for each point P in P OS do
Reproject P to image plane, get coordinates of pixel Compute color of pixel (x, y) for O and t.
(x, y) and signed distance t to P from current camera if Ray(x, y) hits an object O at the point P and the
viewpoint. signed distance t then
for each (ix, iy) in 33 neighborhood of pixel (x, y) Add triple (P, O, t) to P OS at (x, y)
do end if
if (ix, iy) does not lie outside the image then end for
if AIP (ix, iy).O 6= 0 then
if AIP (ix, iy).t t then
AIP (ix, iy).t = t 4.1 View Frustum Error
if AIP (ix, iy).O P.O then The view frustum error is caused by appearance of an ob-
Increment(AIP (ix, iy).cnt) ject that was previously outside of the current view frus-
else tum. Since there is no information about such an object in
{Rewrite by closer object.} the P OSi1 array, it cannot be handled by the reprojec-
AIP (ix, iy).O = P.O tion. This error can occur when the center of projection Ci
AIP (ix, iy).cnt = 1 is outside the view frustum of the previous frame.
end if Fig. 2 shows an object Y that was not visible from C1 ,
else but becomes visible from C2 . The reprojection cannot
if AIP (ix, iy).O P.O then handle the appearance of Y since the object was not in-
Increment(AIP (ix, iy).cnt) tersected by any ray from C1 .
end if
end if
else
{First assignment to AIP (ix, iy)}
AIP (ix, iy).O = P.O
AIP (ix, iy).t = t
AIP (ix, iy).cnt = 1
end if
end if
end for
end for Figure 2: View frustum error. Reprojection cannot han-
dle appearance of an object that was outside of the view
frustum.

4 Reprojection Errors To solve the view frustum error we check if the current
camera position is outside the view frustum of the previous
Reprojecting points from 3D space onto the image plane camera position. In such a case we skip the reprojection
does not guarantee the correctness of results for ray cast- phase and compute visibility of all pixels by the ORSA.
ing. There are two cases when the reprojection can fail: However, since in a walkthrough the camera is typically
the view frustum error and the occlusion error. Similar moving forward or rotates, the reprojection is almost al-
sources of error in the context of image based rendering ways possible. Note that the view frustum error cannot
are called the external and internal exposure errors [19]. occur for objects that appear in the view frustum due to
We discuss these two errors in the context of our ray cast- camera rotation. Such objects are either behind entries in
ing algorithm and suggest methods to handle them. P OSi1 array (if also translation forward is involved) or
there are no entries for the corresponding pixel and thus distance on V Pi for the reprojected point is smaller, the
the ORSA is invoked. point is a possible source of occlusion error and we dis-
card it from the reprojection.
To determine the correct reprojection order we subdi-
4.2 Occlusion Error
vide the image plane V Pi1 into four quadrants according
The occlusion error is caused by appearance of previously to the epipole Ei1. Within each quadrant we apply a dif-
occluded objects. If an occluded object becomes visible at ferent processing order of the points to guarantee that the
a pixel (x, y), it is possible that the reprojection suggests reprojected point has a larger distance from the epipole
that another object is visible at (x, y). This follows from Ei1 than the previously projected points [18][page 62].
the fact that we have no information about the previously The ordering is illustrated at the top of Figure 4.
occluded objects or their occluded parts.
The occlusion error is illustrated in Fig. 3. For the
camera position C1 object Y occludes object X between
points c and d. For the camera position C2 the points on X
between b and c reproject to the new viewing plane. Thus
if there was a previously occluded object in the marked
quadrilateral, it is missed by the reprojection phase. The
reprojection cannot determine in advance whether there is
an occluded object. We apply a method that identifies re-
projected points that interfere with previously occluded re-
gions in the reprojection (between b and c in our example).

Figure 3: Reprojection in general can result in incorrect

results even if a new camera position lies in the view frus-
tum of the previous camera position. The scene consists of
two objects X and Y . If there was an object in the marked
quadrilateral (in yellow color) it could be missed due to
reprojection of points between b and c.

The occlusion problem has been dealt in the context of Figure 4: (Top) The order of point reprojection with re-
image based rendering [19, 18], volumetric rendering [12], spect to the V Pi1. (Bottom) Result of the reprojection
and visibility preprocessing [16]. Our solution to the prob- to V Pi . The order of reprojection and the min/max repre-
lem is an extension of the method of Gudmundsson [12]. sentation of the already projected points identifies possible
We apply a verification of a reprojected point with respect occlusion problems.
to the previously reprojected points that is based on their
depths. If the verification fails we discard the reprojected To check if the distance of a reprojected point from the
point. epipole E2 is increased we keep maximal orthogonal dis-
The verification technique exploits properties of epipo- tances of already reprojected points for each column and
lar geometry of two camera positions. We project the row of AIPi . For each reprojected point we compute its
current camera position Ci to the previous viewing plane vertical distance to the row of the epipole and its horizon-
V Pi1 to obtain the epipole Ei1. Similarly we project tal distance to the column of the epipole. These distances
Ci1 to V Pi to obtain the epipole Ei. The points are are compared with the min/max distances (depending on
reprojected in an order of increasing distance from the the quadrant) stored at the column and the row the point
epipole Ei1 on V Pi1. We check if the points exhibit projects to. If both distances are increased (in absolute
increasing distance from the epipole Ei on V Pi . If the values) we update min/max for the row and the column
and insert the reprojected point to AIPi . Otherwise the distance at each frame. We assume that the object Os that
point is discarded since it can be a source of the occlusion we want to detect is attached to a larger already visible ob-
error. The described processed is illustrated in Figure 4. ject. Thus we compute , taking t as a signed distance of
each reprojected point. is compared with the size of the
4.3 Undersampling Errors camera movement. If the camera movement is larger than
, the point is approaching the camera too fast and it is
The errors discussed above can occur due to missing in- discarded from the reprojection.
formation about invisible scene objects either due to view Setting N = 9 in this context is theoretically correct
frustum restrictions or due to occlusion. In this section we but very conservative. In practice, setting N = 4 was
discuss an additional source of error that can occur due to sufficient and did not result in any visible artifacts in our
the discrete character of the method, namely the insuffi- test walkthroughs.
cient sampling density.
Ray casting computes visibility of scene objects by sam-
pling and so some small objects (below pixel size) can 5 Implementation
be missed. As the camera is moving the objects should
become visible, but since we have no information about In this section we briefly discuss the implementation issues
these objects they cannot be handled by the reprojection. and two optimization techniques.
As a solution to the problem we suggest to use regular
resampling in the spatio-temporal domain. For each N -
th frame we compute some pixels of the image using the 5.1 Implementing Reprojection
ORSA. We use disjoint sets of pixels for each of the N
The reprojection of a point is applied for each entry of the
subsequent images, which guarantees that after N frames
P OS array. Therefore an efficient implementation of the
all the pixels in N images were at least once computed
reprojection is a key to achieve performance improvement
exactly by the ORSA. We used N = 9, thus in each
of the proposed method.
frame we compute one pixel in a window of 3 3 pix-
Unlike reprojection algorithms used in image-based
els using the ORSA (see Figure 5). The sampling pattern
rendering [19, 18], we need to compute a signed dis-
aims to minimize discrepancy in the spatio-temporal do-
tance of the reprojected point from the camera center (re-
main for straightforward camera motion. The pattern has
quired in Algorithm 1). Our implementation of reprojec-
been found experimentally by computing discrepancy of
tion based on the principle component of camera forward
all possible patterns for 3 3 window. ~ precomputes 7 coefficients that are valid for the
vector N
whole image. The reprojection of one 3D point requires
the following operations: 8 (+), 3 (), 12 (), 1

(/), 1 ( ), and 2 (IF ). In our implementation the
reprojection takes time comparable to computation of an
intersection between a ray and a sphere. Thus it is reason-
ably fast not to become a bottleneck of the method.

Figure 5: Regular resampling pattern for a 3 3 window. 5.2 Handling Background

In k-th frame, all pixels denoted by (k mod 9) + 1 are
The reprojection technique as described accelerates ray
computed by the ORSA.
casting only for pixels at which some object is visible. We
can extend our method to handle the empty pixels as fol-
lows: we represent the scene background by a bounding
4.3.1 Abrupt Camera Motion sphere of the scene centered at the center of the camera.
The regular resampling is sufficient for smooth changes The reprojection then handles the sphere in the same way
in camera position. For abrupt forward camera motion as the other scene objects.
some small objects can still be undersampled and missed.
We calculate for each new reprojected point if the forward 5.3 Shifting Ray Origin
movement of the camera was small enough to detect this
object by regular resampling. Even if the reprojection fails due to insufficient num-
Let an object Os at the signed distance t have footprint ber of reprojected points (Section 3.3) we know that the
on the image plane of size S, thus at the signed distance ray/object intersection cannot occur within certain signed
t/2 the footprint of size 2S (the area of the footprint is distance from the viewpoint. We can thus start the ORSA
four times larger). When S is below pixel size, it should with a shifted ray origin. By shifting a ray origin forward
be safely detected within N frames by regular reprojec- along the ray we avoid the traversal of empty nodes of a
tion, if we do not move the camera faster than = t/2/N spatial data structure used for computing the ray traversal.
 The ray origin is shifted from the center of projection Ci RayCasting Parameters
by t , where t is the computed signed distance and  is a Method TR FR IT N
N T S cE [%] cR [%]
small positive constant required for numerical robustness ORSA 701.1 1.17 9.23 52.0
of the algorithm.
REP R1 478.4 1.72 3.37 12.2 10.9 78.0
REP R2 478.3 1.72 3.35 11.8 10.5 78.4
6 Results IRSA 344.3 4.05 0.98 0.0

As the ORSA we use a ray shooting algorithm using a Table 1: Results for the townhouse scene (85k objects).
kd-tree built according to the surface area heuristics with The walkthrough consists of 825 frames. TR is the total
split-clipping [13]. The reprojection algorithm was im- time of rendering (including Phong shading), FR is the
plemented in C++ and compiled using gcc-3.0.3 compiler. average frame rate. ORSA is the ordinary ray shooting
The tests were conducted on a PC equipped with AMD algorithm based on a kd-tree for each pixel. REP R1 is the
Athlon 1.1 GHz, 768 MB RAM, running Linux. reprojection algorithm as described in Section 3. REP R2
We have tested our reprojection algorithm on walk- is REP R1 combined with shifting ray origin. IRSA is
N IT < 1 since
an ideal ray shooting algorithm; note that N
throughs of three different scenes: townhouse, office,
and theatre. The townhouse scene contains 85k objects a background is visible.
Method TR FR IT
N T S
N cE [%] cR [%]
(see snapshot in Fig. 6-left). The office scene contains
626k objects. The theatre scene contains 112k objects. ORSA 979.6 1.19 3.79 58.1
The walkthroughs were rendered in resolution of 512512 REP R1 837.9 1.39 2.73 26.0 31.1 57.8
pixels. REP R2 835.9 1.39 2.69 24.6 31.1 57.8
In order to quantify the success of the reprojection we IRSA 373.8 3.11 0.999 0.0
compute the percentage of pixels computed by the single
ray/object intersection (cR [%]) and by the ORSA due to Table 2: Results for the theatre scene (112k objects). The
failure of the reprojection (cE [%]). The number of pixels walkthrough consists of 1165 frames.
computed by the ORSA due to regular resampling was Method TR FR IT
N T S
N cE [%] cR [%]
constant for all walkthroughs (cO = 11.1%). The results ORSA 892.3 1.66 3.02 44.5
show the performance of the ORSA, two variants of our REP R1 789.4 1.89 1.89 21.9 33.6 55.3
reprojection method (REP R1 and REP R2), and an ideal REP R2 801.5 1.85 1.88 21.7 33.6 55.3
ray shooting algorithm IRSA. The ideal ray shooting
IRSA 354.6 4.19 0.71 0.0
algorithm[13] is an abstract concept that uses the results
of the ORSA to determine the performance of computing Table 3: Results for the theatre scene (626k objects). The
ray/object intersections if the intersected object is already walkthrough consists of 1489 frames.
known. The results are summarized in Table 1, 2, and 3.
The proposed method achieves the following increase
of the average frame rates: 1.17 to 1.72 for the townhouse
scene, 1.66 to 1.89 for the office scene, 1.19 to 1.39 for the ideal ray shooting algorithm IRSA. These numbers
the theatre scene. We note that the ORSA used in our test include time of shading at the computed ray/object inter-
is an already highly optimized algorithm that is considered section points. If we exclude time for shading and eval-
one of the most efficient ray shooting techniques [29]. The uate speedups for pure visibility computation, we achieve
reprojection algorithm would provide a more significant frame rate increases for computing visibility between 18%
speedup for less efficient ray shooting algorithms. and 82%. This evaluation is based on code profiling; the
profiler reported that about 30% of the total rendering time
of the ORSA was spent in the shading routine.
7 Discussion The speedup obtained by the reprojection is propor-
tional to the number of pixels the reprojection succeeds
7.1 Visual Quality on. The utilization of reprojection increases with the reso-
Firstly, we verified the visual quality of the generated im- lution of generated images and decreases with the number
ages. As a reference we used images generated by the of visible silhouettes and their length. The length of the
ORSA on the same walkthrough sequence. We did not silhouettes in the image increases linearly with resolution,
observe any visual artifacts caused by the reprojection for but the number of pixels rendered increases quadratically.
all tested walkthroughs. This suggests that the presented method has a greater po-
tential for a high resolution rendering.
Surprisingly we have observed that the acceleration
7.2 Speedup
technique that shifts the ray origin (REP R2) did not per-
We increased the frame rates from 14% to 47% and form as well as expected. It provided a slight speedup
achieved the performance 1.70 to 2.23 times slower than for the theatre scene, but for the office scene it led to a
slight performance loss. We conclude that this is caused and proposed methods that handle these errors. In par-
by its computational overhead that is simply too high to ticular we used splatting, checking camera position with
further improve the already quite efficient ray casting im- respect to the view frustum, ordered reprojection, regular
plementation. The technique for handling the background resampling, and handling abrupt camera motion. We sug-
is computationally inexpensive, but since the tested scenes gested two optimization techniques to handle the image
did not contain many empty pixels its impact could not be background and to accelerate ray traversal by shifting the
properly evaluated. ray origin.
We verified our algorithm by a series of tests. On the
tested walkthrough sequences the reprojection succeeded
7.3 Limits of Ray Casting Acceleration in determining of up to 78% pixels. The frame rate was
increased approximately by 14 to 47% for moderately
We discuss the limits of ray casting acceleration and relate
sized scenes. We discussed the achieved results in the
the speedup provided by our technique to these limits. Our
context of the limits of ray shooting acceleration. We
discussion is based on the abstract concept of the ideal ray
have shown that the proposed technique is relatively close
shooting algorithm (IRSA) that knows exactly which ob-
to the ideal ray shooting algorithm. In particular it is
ject is intersected by the given ray. In other words such an
1.70 to 2.23 times slower than the ideal state. This obser-
algorithm acts as an ideal predictor for each input ray.
vation suggests that a further purely algorithmic accelera-
The experimental measurements (Tables 1, 2, and 3)
tion of ray casting with conservative results is very com-
show that even the ORSA based on kd-trees is rela-
plicated. We see a greater potential for acceleration by
tively close to the optimal algorithm: it is 1.93 to 2.75
reducing the number of rays cast or by explicit paralleliza-
times slower than IRSA for the tested scenes. Also the
tion.
hardware-independent values such as number of ray/object
In the future we plan to investigate properties of our al-
intersections and the number of traversal steps are very
gorithm in context of global illumination methods. We
close to ideal state. Thus the space for an improvement
would also like to extend the algorithm for scenes with
is already quite small. The main reason why the perfor-
moving objects.
mance increase of our method is relatively small (14% to
47%) is the proximity of the achieved running time to the
IRSA. Judging from these results we conjecture that fur- Acknowledgements
ther algorithmic acceleration of ray shooting is very com-
plicated. We see a greater potential for acceleration by re- Here, we want to thank Philippe Bekaert for providing
ducing the number of rays cast [23] (thus obtaining less us the scene models for research purposes and Jaroslav
precise images) or by exploiting explicitly parallel com- Krivanek for having comments on previous version of the
putations [29]. paper. Further, we would like to thank all anonymous re-
viewers of the previous version of the paper. This work has
been partially supported by the Austrian Science Founda-
7.4 Extension for Shadow Rays tion (FWF) contract no. p-13876-INF and by IST 2001-
In the scope of ray tracing it is significantly more difficult 34744 funding.
to exploit temporal coherence of secondary rays than the
coherence of primary rays [1]. For a moving camera the
secondary rays have low coherence with respect to direc-
References
tions and origins of the rays in the previous frame. How- [1] S.J. Adelson and L.F. Hodges, Generating exact
ever, the reprojection can be extended to handle direct il- ray-traced animation frames by reprojection, j-
lumination by point-light sources similarly to method of IEEE-CGA, 15(3):4352, 1995.
Adelson [1]. This technique can also be extended to han-
[2] A. Appel, Some techniques for shading machine
dle area light sources approximately.
renderings of solids, in AFIPS 1968 Spring Joint
Computer Conf., vol. 32, pp. 3745, 1968.
[3] J. Arvo and D. Kirk, A survey of ray tracing accel-
8 Conclusion and Future Work eration techniques, pp. 201262. Academic Press,
We presented a novel method for ray casting acceleration 1989.
that exploits temporal coherence in the scope of walk- [4] J.S. Badt, Two algorithms for taking advantage
throughs of static scenes. The proposed method uses re- of temporal coherence in ray tracing, The Visual
projection of the set of visible points from the previous Computer, 4(3), pp. 123132, 1988.
frame to guide ray casting in the current frame. The aim [5] J. Bittner and V. Havran, Exploiting temporal and
of our technique was to eliminate the ray traversal and to spatial coherence in hierarchical visibility algo-
compute only a single ray/object intersection per pixel. We rithms, in Journal of Visualization & Computer
discussed several source of errors due to the reprojection Animation, 12(5):277286, 2001.
 (a) (b)

Figure 6: (a) A snapshot from a walkthrough of the townhouse scene. (b) Visualization of the reprojection and the ren-
dering phases. Red pixels are computed successfully by a single ray/object intersection, green pixels have been computed
exactly by the ORSA due to reprojection failure, white pixels are computed by the ORSA due to regular resampling, and
blue pixels are computed by the ORSA due to detection of abrupt camera motion. The profiler reported 14.7% of time
spent by reprojection phase, 41.4% spent by computing the normal and Phong shading, 2% spent by the initialization of
rays, 25.5% spent by the ORSA, 3.24% spent by computing single ray/object intersection, and the rest (13.2%) spent in
the main ray casting function.

[6] Catmull, E.E. (1975). Computer display of curved ceedings in The First Conference on Visualization
surfaces. In Proceedings of the IEEE Conference on in Biomedical Computing, Atlanta, 1990.
Computer Graphics, Pattern Recognition, and Data
Structure, 1117. [13] V. Havran, Heuristic Ray Shooting Algorithms. PhD
thesis, Czech Technical University in Prague, Nov.
[7] S. Coorg and S. Teller, Temporally coherent con- 2000.
servative visibility, in Proceedings of the Twelfth
Annual Symposium On Computational Geometry [14] H. Hey and W. Purgathofer, Occlusion culling
(ISG 96), pp. 7887, ACM Press, May 1996. methods, in STAR in Eurographics 2001, Euro-
[8] S. Coorg and S. Teller, A spatially and temporally graphics Association, Sept. 2001.
coherent object space visibility algorithm, Tech. [15] Q. Lin, P.-E. Danielson, and B. Gudmunson,
Rep. TM-546, Department of Computer Graphics, Frame-coherent volume rendering, in Proceed-
MIT, Feb. 1996. ings of the IASTED International Conferent, COM-
[9] J. Demers, I. Yoon, T.-Y. Kim, and U. Neumann, PUTER GRAPHICS AND IMAGING, Nov. 2000.
Accelerating ray tracing by exploiting frame-
to-frame coherence, Technical Report 1998/668, [16] X. Liu, H. Sun, and E. Wu, Visibility preprocess-
Computer Science Department, The University of ing with occluder fusion for urban walkthroughs,
Southern California, Los Angeles, 1998. Proceedings of VRST2000, pp. 55-60, 2000.
[10] F. Durand, 3D Visibility: Analytical Study and Ap- [17] W. Martin, S. Parker, P. Shirley, and W. Thomp-
plications. PhD thesis, Universite Joseph Fourier, son, Temporally coherent interactive ray tracing,
Grenoble, France, July 1999. Tech. Rep. UUCS-01-005, Computer Science De-
[11] N. Greene and M. Kass, Hierarchical Z-buffer vis- partment, University of Utah, May 2001.
ibility, in Proc. of SIGGRAPH93, pp. 231240, [18] W.R. Mark, Post-Rendering 3D Image Warping:
1993. Visibility, Reconstruction, and Performance for
[12] B. Gudmundsson and M. Randen, Incremental Depth-Image Warping. PhD thesis, Chapel Hill,
generation of projections of ct-volumes., in In Pro- University of North Carolina, 1999.
[19] L. McMillan, An Image-Based Approach to Three-
Dimensional Computer Graphics. PhD thesis,
Chapel Hill, University of North Carolina, 1997.
[20] H. Pfister, M. Zwicker, J. van Baar, and M. Gross,
Surfels: Surface elements as rendering primi-
tives, in Siggraph 2000, Computer Graphics Pro-
ceedings, pp. 335342, 2000.
[21] E. Reinhard, P. Shirley, and C. Hansen, Paral-
lel point reprojection, in Proceedings of the 2001
IEEE Symposium on Parallel and Large-Data Visu-
alization and Graphics, Oct. 2001.
[22] S. Rusinkiewicz and M. Levoy, QSplat: A
multiresolution point rendering system for large
meshes, in Siggraph 2000, Computer Graphics
Proceedings, pp. 343352, 2000.
[23] A. Scheel, M. Stamminger, J. Putz, and H.-P.
Seidel, Enhancements to Directional Coherence
Maps, in Proceedings of WSCG01, pp. 403410,
2001.
[24] M. Simmons, A dynamic mesh display representa-
tion for the holodeck ray cache system, Tech. Rep.
CSD-00-1090, University of California, Berkeley,
2000.
[25] M. Simmons, C.H. Sequin, Tapestry: A Dynamic
Mesh-based Display Representation for Interactive
Rendering, in Rendering Techniques 00, pp. 329-
340, 2000.
[26] O. Sudarsky, Exploiting temporal coherence in
animation rendering, Technical Report CIS9326,
Technion - Computer Science Department, Israel,
1993.
[27] L. Szirmay-Kalos and G. Marton. Worst-case ver-
sus average case complexity of ray-shooting. Com-
puting, 61(2):103131, 1998.
[28] S. Teller and J. Allex, Frustum casting for progres-
sive, interactive rendering, Tech. Rep. MIT LCS
TR-740, MIT, Jan. 1998.
[29] I. Wald, P. Slusallek, C. Benthin, and M. Wagner,
Interactive rendering with coherent ray tracing,
in EG 2001 Proceedings, vol. 20(3) of Computer
Graphics Forum, pp. 153164, 2001.
[30] B. Walter, G. Drettakis, and S. Parker, Interactive
rendering using render cache, in Rendering Tech-
niques 99, pp. 1930, 1999.
[31] R. Yagel and Z. Shi, Accelerating volume anima-
tion by space-leaping, in Proceedings of Visualiza-
tion93, pp. 6269, Oct. 1993.