causal - illusiory motion casual time filtering.pdf

ng

omp

mpo. Wxate syto etricsorwa

y strononguripatterric inteddle rehe Japmatic(the b

the dark bar to dark gray) (Kitaoka & Ashida, 2003).The illusory movement is experienced under free viewing con-

ditions when one moves the eyes, and it is perceived in non-centralvision. It stops if steadily xating after about 68 s. The perceivedmotion is a drift, whose direction depends on the intensity rela-tionship of the pattern elements. Chromaticity is not necessaryfor the illusion, but enhances the effect in some patterns (Backus& Oru, 2005; Kitaoka, 2006). The illusion depends on the size of

the image patterns. For medium sized patches such as Donguri,

intensity proles giving rise to the illusory effect.A number of hypotheses for the illusory motion have been pro-

posed. The dominant idea originating from Faubert and Herbert(1999) is that temporal differences in luminance processing pro-duce a signal that tricks the motion system. The theories differ inhow this signal is produced. Faubert and Herbert suggest thateye movements or blinks need to trigger an image motion, andthe different motion signals (due to differences in intensity) areintegrated over large spatial areas. Backus and Oru (2005) focuson the perception during xation and hypothesize that motion isnot necessary, but a motion signal is triggered from the change

* Corresponding author. Fax: +1 301 414 9115.

Vision Research xxx (2009) xxxxxx

Contents lists availab

Re

.e l

ARTICLE IN PRESSE-mail address: [email protected] (C. Fermller).next to different shades of gray. Referring to Fig. 3b, from the high-est intensity (the white bar) the intensity drops about twice asmuch on the right than on the left side. Similarly, from the lowestintensity (the dark bar) the intensity rises about twice as much onthe right than on the left. Thus, at the two bars the change of inten-sity in the right and left neighborhood is different. Informally wesay that the pattern is asymmetric. Patterns with such intensityproles create a very strong illusory effect. The perceived move-ment is a drift from the intensity extremum in the direction of les-ser intensity change (i.e. from the white bar to light gray, and from

grating named the staircase illusion (Fraser & Wilcox, 1979) andthe peripheral drift illusion (Faubert & Herbert, 1999) (seeFig. 18). Ashida and Kitaoka (2003) showed that the effect is muchincreased if the sawtooth luminance prole is replaced by stepfunctions with intensities in the same order as in Donguri (i.e. lightgraywhitedark grayblack), and if the large patches are replacedby many small ones. In Kitaoka and Ashida (2004) the authors pre-sented patterns with continuously increasing intensity ramp-likeproles, which are perceived in central and close to central vision,and Kitaoka in (2006) proposed a classication of the different1. Introduction

Most observers experience verwhen viewing patterns such as DSnakes (Fig. 2) (Kitaoka, 2003). Theseage patches which have an asymmetple, consider a narrow slice in the miin Donguri, as shown in Fig. 3a. (Ttranslates to acorn.) Its monochrodescribed as a white and a dark bar0042-6989/$ - see front matter 2009 Elsevier Ltd. Adoi:10.1016/j.visres.2009.11.021

Please cite this article in press as: Fermller, C.,g illusory movement(Fig. 1) and Rotatingns are composed of im-nsity prole. For exam-gion of one of the ovalsanese word donguriintensity image can beoundaries of the oval)

motion occurs in a patch when it is viewed in the periphery. Smal-ler patches give illusory motion closer to the center of the retina.Blur reduces the illusion in peripheral vision, but enhances it incentral vision. The illusory effect is more forceful if a pattern con-sists of many patches, and the patches are at multiple sizes. It isstronger when the patches are circularly organized, but also existsfor columnar and other arrangements.

It is generally considered that the illusory effect was rst ob-served in patterns with circularly organized sawtooth luminanceIllusory motion due to causal time lteri

Cornelia Fermller a,*, Hui Ji b, Akiyoshi Kitaoka c

aComputer Vision Laboratory, Center for Automation Research, Institute for Advanced CbDepartment of Mathematics, National University of Singapore, SingaporecDepartment of Psychology, Ritsumeikan University, Japan

a r t i c l e i n f o

Article history:Received 8 April 2008Received in revised form 24 November 2009Available online xxxx

Keywords:Illusory motionSpatio-temporal lteringImage motion estimateComputational model

a b s t r a c t

A new class of patterns, coception of illusory motionimage motion induced byral energy lters, which arbut not the future, is usedmotion of locally asymmeception of the different illuopposing real motion, and

Vision

journal homepage: wwwll rights reserved.

et al. Illusory motion due to cauter Studies, University of Maryland, College Park, MD 20742-3275, United States

sed of repeating patches of asymmetric intensity prole, elicit strong per-e propose that the main cause of this illusion is erroneous estimation ofional eye movements. Image motion is estimated with spatial and tempo-mmetric in space, but asymmetric (causal) in time. That is, only the past,stimate the temporal energy. It is shown that such lters mis-estimate theintensity signals at certain spatial frequencies. In an experiment the per-y signals was quantitatively compared by nulling the illusory motion withs found to be predicted well by the model.

2009 Elsevier Ltd. All rights reserved.

le at ScienceDirect

search

sevier .com/locate /v isresusal time ltering. Vision Research (2009), doi:10.1016/j.visres.2009.11.021

2 C. Fermller et al. / Vision Rese

ARTICLE IN PRESSof the neural response over time. Differently strong contrasts andintensities cause different neural response curves over time(Albrecht, Geisler, Frazor, & Crane, 2002). As a result the phase ofthe signal is estimated erroneously as time passes and a motionsignal is triggered. Their model also introduces the effect of adap-tation which can account for the smooth perception under xationover a few seconds. This effect may exist in addition to the one dis-cussed here. Conway, Kitaoka, Yazdanbakhsh, Pack, and Living-stone (2005) discuss the illusory motion effect when ashing the

Fig. 1. Variation of Donguri pattern. In peripheral vision most observers experiencerotary movement. The direction in the circular arrangements alternates, withcounterclockwise direction in the upper left.pattern. Their viewpoint is that small eye movements refreshretinal stimulation, promoting new onset responses in the stimu-lated area on the retina. Thus the system has available a stimulusconsisting of pattern frames interlaced with frames of the patternsaverage intensity. Using psychophysical and physiological experi-ments Conway et al. (2005) argue that in addition to signals cre-ated by the differences in intensity processing, a signal analogousto the reverse phi motion (Anstis, 1970) is created. Reverse phi mo-tion is an image motion effect caused by reversing the contrast insome frames of a video sequence. It is easy to explain that theashing illusion will produce an apparent motion (phi motion) inthe direction of the perceived motion. On appearance of the pat-tern the net motion of the pattern is in the direction described be-fore and on disappearance it is in the opposite direction. In ouropinion the illusory effect under free viewing and the effect whenickering the pattern are not the same. We observed that for thereduced experimental stimuli the latter is experienced much stron-ger than the former, and it is experienced even by observers whodo not have the effect under free viewing. We therefore do not be-lieve that Conway et al. provide a sufcient account of the illusion.

We propose that the main reason for the illusion under freeviewing conditions is erroneous estimation of the image motiondue to involuntary xational eye movements. Work by Murakami,Kitaoka, and Ashida (2006) implicates drift eye movements.1 In

1 The drift movements, one of the three xational eye movements, are dened asincessant random uctuations at about 130 Hz, quite large (10 min of visual angle)and fast (up to 23/s) (Eizenman, Hallett, & Frecker, 1985). They have greateramplitude after a saccade (Ross, Morrone, Goldberg, & Burr, 2001) than during steadyxation.

Please cite this article in press as: Fermller, C., et al. Illusory motion due to caparticular, the authors showed a correlation between the ampli-tude of drift movements in different observers and the strengthof their illusory perception. Further evidence for the role of driftsin this illusion comes from the fMRI studies of Kuriki, Ashida,Murakami, and Kitaoka (2008). Comparing the snake illusion witha control stimulus, they found signicantly increased activity inmotion area MT+ (also called V5) when eye movements were pres-ent, but no increase in the absence of eye movements.

The small eye movements cause a change of the image on theretina and trigger the estimation of a motion eld. This motioneld is due to rigid motion and thus has a certain structure. Undernormal circumstances the vision system estimates this image mo-tion and compensates for it, i.e. the images are stabilized (Muraka-mi, 2004; Murakami & Cavanagh, 1998). Even for asymmetricsignals, the vision system estimates the correct 3D rigid motionusing the average of all the motion vectors in the patterns. How-ever, a mis-estimation occurs at certain locations in the image.The difference between the estimated rigid motion eld and theerroneously estimated image motion vectors gives rise to residualmotion vectors. These residual motion vectors are integrated overtime and space causing the perception of illusory motion in theimage.

The dominant model for motion processing in humans andother mammals is the motion energy model (Watson & Ahumada,1985; Adelson & Bergen, 1985), and it has been found to be consis-tent with the physiological responses in primary visual cortex(Albrecht & Geisler, 1991). Motion is found from the response ofmultiple spatio-temporal lters, which are separable in spaceand time. The spatial lters are symmetric. The temporal lters,however, are asymmetric. This is because real-time systems havecausal lters, which are lters that receive as input data from thepresent and the past, but not the future. If such lters were sym-metric, the processing would be delayed by half the extent of thelter. Since early responding is valuable, the temporal responsesare asymmetric in time, with greater weight given to recent inputthan older input.

As will be shown, causal lters mis-estimate the image motionin asymmetric image signals for certain spatial frequencies. That is,if we apply differently sized motion lters to some asymmetricpattern, we will get mis-estimation for a range of lter sizes. Theresolution of the eye decreases from the center to the periphery.Thus, the size of the motion energy lters increases as we movefrom the center to the periphery, and their spatial frequency de-creases. The illusory motion patterns consist of repeated patchesof asymmetric signals, and for some of these patches the resolutionof the eye is such that it leads to erroneous motion. For most of theknown patterns the mis-estimation occurs at the periphery. Forvery small patterns with high frequency the perception is closerto the center.

The next two sections will explain in detail the reasons for themis-estimation of image motion. The reader not interested in thetechnical details may want to skip these sections. We summarizehere the main concept: Fig. 4 illustrates a spatio-temporal lterwith symmetric impulse response in the spatial domain, and withasymmetric impulse response (Burr & Morrone, 1993) in the timedomain. The spatial response may be modeled as a sinusoid ofcertain frequency enveloped by a symmetric function. The tempo-ral response may be modeled as a sinusoid enveloped by an asym-metric function. Consider ltering the Donguri signal with awhole range of spatial lters of increasing size (and decreasingfrequency). Let us go ahead in the paper and take a look atFig. 8bd, which show the amplitude of the response from lter-ing a single bar in Donguri. A lter of high spatial frequency will

arch xxx (2009) xxxxxxrespond to the two edges bordering the bar. A lter of low spatialfrequency will not recognize the edges, but only have one re-sponse to the bar. However, for intermediate frequencies, with

usal time ltering. Vision Research (2009), doi:10.1016/j.visres.2009.11.021

the period of the sinusoid about as large as the bar, the two edgeswill effect each other during ltering leading to an amplitude re-sponse curve of a larger peak merged with a smaller peak. In es-sence, for lters of these frequencies there is poor frequencylocalization. While the ltered signal should have the frequency

If now we estimate image motion by applying to this signal tem-poral asymmetric temporal lters, we nd that the temporal fre-quency responses from a movement to the left and a movementto the right will be signicantly different. The image motion esti-mated as the average over the signal is larger for left motion than

Fig. 2. Rotating snakes.

C. Fermller et al. / Vision Research xxx (2009) xxxxxx 3

ARTICLE IN PRESSof the lter everywhere, the actual value varies along the signal.Fig. 3. (a) Slice through a patch in the Donguri pattern. (b) Its intensity prole. (c and d) Sa static image. I0t and I0t denote the signal at a point over time, where the sign in Iobtain the inverted motion direction.

Please cite this article in press as: Fermller, C., et al. Illusory motion due to cafor right motion.patio-temporal picture of the patch moving to the left and to the right. I0x denotes0t indicates that the prole in (d) can be obtained by reecting the prole in (c) to


quency space. We follow the common formulation of modeling a

Rese

ARTICLE IN PRESSThe idea of anisotropic temporal ltering was rst proposed byAshida and Kitaoka (2003), who modeled image motion estimationusing a differential local motion model with asymmetric temporalderivative lters. This local model, however, requires the lters tohave larger weight in the past and smaller weight in the present.Furthermore, it cannot explain the estimation at different resolu-tions of the pattern. For this we need to look at the differentfrequencies.

2. Motion estimation in the frequency domain

The monochromatic light distribution on the retina can be de-scribed as a function Ix; y; t, which species the intensity at apoint x; y at time t. We refer to the instantaneous light distribu-tion at time t 0 as the static image I0x; y Ix; y;0. Let us as-sume that within a small interval the change of the image can bedescribed as a translation with constant motion velocity ~v of hor-izontal and vertical speed components u; v. Thus, the intensityfunctions at time t and at time 0 are related as

Ix; y; t Ix ut; y vt;0: 1From the three-dimensional Fourier transform of this equation, weobtain (Watson & Ahumada, 1985; Adelson & Bergen, 1985)

Fig. 4. Illustration of biological implementation of spatio-temporal lter (similar toFig. 6 of Adelson and Bergen (1985)). The spatial and temporal impulse responsesare shown along the margins. Their product is shown schematically in the center.

4 C. Fermller et al. / Visionuxx vxy xt : 2where xx; xy denote the spatial frequencies and xt the temporalfrequency. This equation denes a plane through the origin in thethree-dimensional frequency space.

To simplify the analysis, in the following sections we consideronly images with bar-like structures parallel to the vertical dimen-sion and the motion component perpendicular to the bars. Thus, letus consider a two-dimensional case of Ix; t, that is a signal Ixwhich is shifted. Eq. (1) then simplies to

Ix; t Ix ut;0; 3and the image motion constraint amounts to

uxx xt; 4dening a line in the two-dimensional frequency space. The velocityu can be found from the ratio of the temporal and spatial frequency,i.e. as

u xtxx

: 5

Fig. 3c and d illustrate the spatio-temporal signal for an image linein the Donguri pattern moving with velocity u 1 and u 1,

plausibly chosen as rx 1 . The transfer function of the Gabor lter,

Please cite this article in press as: Fermller, C., et al. Illusory motion due to caxxwhich is obtained as its Fourier spectrum, amounts to G^x;xx exp2p2r2x xxx2, that is a Gaussian centered at xx and ofstandard deviation 12prx. The Gabor of frequency xx, thus extractsthe signals energy in a small frequency band around xx. Fig. 5billustrates the amplitude of G^. Its phase is zero (i.e. there is no imag-inary part), because the envelope of the Gabor is symmetric around0. In general, symmetric lters around a point different from 0 (forexample, a time-shifted Gabor) have a phase response that is line-arly related to the frequency.

The temporal lter has an envelope described by a functionwith rst-order exponential decay. We use the formulation pro-lter for detecting the local frequency x0, as a complex function

gy py exp2pix0y: 6

exp2pix0y cos2px0y i sin2px0y, called the carrier func-tion, is a complex sinusoidal for detecting the signals componentof frequency x0, and py, called the envelope function, localizesthe sinusoid in image space. The complex lter really consists oftwo lters in quadrature, the even cosine components and odd sinecomponents. For example, in the spatial domain, the even compo-nent will respond maximally to bar-like signals and the odd compo-nent will respond maximally to edges. The magnitude of the outputof the combined complex lter does not depend on whether the sig-nal is even or odd, or any mixture thereof. As a result, complex mo-tion lters (Adelson & Bergen, 1985; Watson & Ahumada, 1985)extract motion independent of the phase of the signal, that is inde-pendent of the position of the signal within the receptive eld atcertain time, and independent of the sign of the contrast.

3.1. Modeling the spatial and temporal lters

We model the spatial lters as Gabor functions (see Fig. 5a)with impulse response

Gx;xx 12p

prx

expx22r2x

exp2pixxx; 7

where the envelope is a Gaussian.xx is the preferred frequency andrx determines the support of the lter, which for convenience isrespectively. Since the spatio-temporal signal Ix; t is obtained sim-ply by shifting the signal I0x, it has the same structure in the spa-tial and temporal domain. Referring to Fig. 3c and d, a spatial cross-section through Ix; t gives a shifted version of I0x. A temporalcross-section gives a shifted, stretched and maybe reected versionof I0. For unit motion to the left u 1, the cross-section is ashifted signal I0t, and for unit motion to the right u 1, it isthe shifted signal I0t (i.e. the reection of I0t). The amount ofstretch encodes the velocity. Thus, later when we analyze temporalltering, instead of examining the temporal cross-section, we canlook at the spatial cross-section.

3. The lters

The spatio-temporal energy lters for extracting motion areseparable in space and time. This just means that the lters canbe created as the product between a spatial and a temporal lter.For the analysis this means that the spatio-temporal signal mayrst be convolved with the spatial lter and the result may thenbe convolved with the temporal lter.

The lters need to be localized in image space as well as in fre-

arch xxx (2009) xxxxxxposed in (Chen, Wang, & Qian, 2001; Shi, Tsang, & Au, 2004), whichmodels the envelope as a Gamma probability density function ofparameter C2, resulting in temporal lters Tt of the form


e reder


ARTICLE IN PRESSTt;xt ts2 exp ts exp2pixtt i/t for t P 00 for t < 0

8

where xt is the temporal frequency. s, the decay velocity, is a timeconstant for the envelope, which we chose as 14xt to make the waveof the temporal lter similar to the Gabor. /t is a phase offset of thesinusoid, which is chosen such that the odd components of the ltersum to zero.

Fig. 6a illustrates the impulse response of this lter. Since thetemporal lter extends from the past to the present, it actuallyestimates the frequency in the recent past. This can also be seenfrom the spectrum of the lter. As can be observed from Fig. 6band c, the amplitude of T^t is still a hat-type function, centeredat xt similar to the Gabor. However, the phase of T^t is non-zero,indicating a shift of the response in image domain. It is approxi-mately linear for x close to xt and deviates from linearity for val-ues farther from xt .

3.2. Denition of ltering

When analyzing image motion, we can think of the ltering as aspatial ltering followed by a temporal ltering. First, the imagesequence Ix; t is ltered with a spatial Gabor, Gx;xx to obtainthe image sequence ~Ix; t;xx as

Fig. 5. Gabor lter of xx 1: (a) impulse response. The full (blue) line denotes thspectrum. (For interpretation of the references to color in this gure legend, the rea~Ix; t;xx Z 11

Iy; tGx y;xxdy:

The idea is that the Gabor obtains the signals component of fre-quency xx. Thus, at this stage it is assumed that the dominant spa-

Fig. 6. Temporal lter of xt 1: (a) impulse response. The full (blue) line denotes the respectrum. (For interpretation of the references to color in this gure legend, the reader

Please cite this article in press as: Fermller, C., et al. Illusory motion due to catial frequency of ~Ix; t;xx at every point x; t is xx. Second,~Ix; t;xx is ltered with the temporal lter Tt;xt to obtainIx; t;xx;xt as

Ix; t;xx;xt

Z 01

~Ix; s;xxTt s;xtds:

Here it is assumed that j Ix; t;xx;xtj2 returns the motion energy of

I at image point x; t at frequencies xx; xt.Since the spatial and temporal lters are complex valued, the

complete spatio-temporal lter can be imagined as four separablelters (the even and odd components of each, the spatial and tem-poral lter), whose outputs are summed according to the rules ofcomplex numbers to arrive at the motion energy.

Fig. 7 illustrates the ltering on the Donguri signal. The spatialand temporal frequencies in this example are set to one (the criti-cal frequencies, to be explained later). Notice, that the temporallter output is shifted with respect to the signal; to the right for leftmotion and to the left for right motion.

3.3. Image motion estimation

In the following we will analyze motion estimation as a func-

al (even) part and the dashed (red) line the imaginary (odd) part. (b) Amplitudeis referred to the web version of this article.)tion of spatial frequency. The image motion of a patch (or in theanalysis a line through the patch), is computed from all the mea-surements in the patch in two computational steps: rst, we esti-mate at every point the (best) velocity. Second, we compute thevelocity of the patch as the weighted average of point-wise velocityestimates.

al part, the dashed (red) line the imaginary part. (b) Amplitude spectrum. (c) Phaseis referred to the web version of this article.)


Rese

ARTICLE IN PRESS6 C. Fermller et al. / VisionSpecically, given a spatial frequency xx, the signal Ix; t is l-tered with the spatial Gabor to obtain ~Ix; t;xx. Then at some timet (all t are equivalent, since the motion is simply a shift of the spa-tial signal) at every image point x, we nd the dominant temporalfrequency xt0 . To do so we lter with a range of temporal ltersTt;xt of different frequencies xt and choose the lter responsewith maximum energy. (Ideally according to the motion con-straint, only one temporal frequency lter should return non-zeroenergy.) This way, we nd at every point x a local velocity estimate

u^x;xx xt0xx 9

and its corresponding energy

j Ix; t;xx;xt0 j2:

Then the motion of a patch is found as the average of energyweighted velocity measurements:

u^xx P

xu^x;xxj Ix; t;xx;xt0 j2P

xj Ix; t;xx;xt0 j2

: 10

4. The effect of ltering on Donguri

Fig. 8 illustrates the effect of spatial ltering on Donguri. At fre-quencies xx larger than the reciprocal width of the bar, the Gaborlter detects the two edges at the left and right of the bar (Fig. 8b).

Fig. 7. Filtering of the Donguri signal at the locations of the horizontal and vertical crosfrequency 1. (b and c) Temporal ltering the signal in (a) with a frequency of 1 for left

Please cite this article in press as: Fermller, C., et al. Illusory motion due to caarch xxx (2009) xxxxxxAt frequencies signicantly smaller than the reciprocal width ofthe bar, the Gabor detects the bar (Fig. 8c). The amplitude of the re-sponse thus has either one or two well separated peaks. However,for frequencies of xx close to the reciprocal width of the bar, thereis something in between one and two responses. The amplitudefunction becomes asymmetric with two merging peaks, a largeron the right and a smaller on the left (Fig. 8d). Let us call these fre-quencies the critical frequencies.

As is well known from the uncertainty principle, there is a limiton the accuracy of localization in image and frequency domain. TheGabor (which is the lter with best localization in joint image andfrequency space) cannot guarantee perfect localization of the sig-nal. Because of the hat prole of its Gaussian envelope, the l-tered signal ~Ix; t;xx will not always have local dominantfrequency xx. We can understand the poor localization of frequen-cies from the phase responses. Referring to Fig. 8eg, the phase re-sponses are (nearly) linear for x 12 and x 2, but the phaseresponse deviates signicantly from linearity for the critical fre-quencies, which is an indicator for poorly estimated frequencies.

When now estimating on the asymmetric signal~Ix; t;xx imagemotion with asymmetric temporal lters, left and right motion areestimated of different value. Fig. 9 shows that for the critical fre-quencies, motion to the left u 1 leads to larger velocity esti-mates than motion to the right u 1.

We can intuitively understand this estimation from the ampli-tudes of the signal and the lter. We convolve two asymmetricsignals. The temporal lter amplitude has more weight for largert and smaller weight for smaller t. Referring to Fig. 7b and c, for

s-sections in Fig. 3c and d: (a) Spatial ltering the intensity signal with a Gabor ofand right motion.


Rese

ARTICLE IN PRESSC. Fermller et al. / Visiona left motion, signal ~I has a larger lobe for larger t and smaller lobefor smaller t, but for right motion the order is reversed.

Fig. 9 shows the local estimated velocity (as full, green line) atevery point on the bar. The corresponding amplitude is shown asdot-dashed, red line, and the amplitude of signal ~I is shown asdashed, blue line (in the spatial domain). Both amplitudes havebeen scaled to allow for better visualization.

Because of interaction of the regions under the two peaks witheach other during temporal ltering, the local velocity (Eq. (9)) var-ies signicantly along the signal. Most signicant, there is overes-timation of velocity at the right peak for left motion, andunderestimation of temporal energy at the left peak for rightmotion.

This is further demonstrated in Fig. 10, which shows the esti-mated energy for three different temporal frequencies. As a result

3 2 1 0 1 2 30

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.1

3 2 1 00

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.1

1 0.5 0 0.5 1

5

0

5

10

15

20

25

30

1 0.5

10

15

Fig. 8. (a) Bar in Donguri pattern. (bd) Amplitude of bar ltered with Gabor of different fof the phase (shown on the y-axis), since it depends linearly on the frequency has been

Please cite this article in press as: Fermller, C., et al. Illusory motion due to caarch xxx (2009) xxxxxx 7of this local mis-estimation, the average velocity (Eq. (10)) is largerfor left motion than for right motion.

For higher spatial frequencies (Fig. 8b) the peaks are well sepa-rated and do not interact, and for lower spatial frequencies (Fig. 8c)there is only one peak. Thus, in both cases there is no signicantdifference between left and right motion.

Two nal notes: Throughout the demonstration we have used anormalized speed of 1 unit, but the ndings apply to any velocity. Adifferent velocity, of say value a, amounts to stretching/compress-ing the signal ~It to ~I ta

. Then a temporal response of xt for the

unit velocity will correspond to a temporal frequency response ofaxt in the stretched signal. Thus, all velocities will be mis-esti-mated by the same percentage.

The size of the motion lter (with non-vanishing energy) in ourimplementation is about ve times the bar width. (The spatial

1 2 3 3 2 1 0 1 2 30

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.1

0 0.5 1 1 0.5 0 0.5 10

2

4

6

8

10

12

14

16

18

20

requenciesxx . (eg) Corresponding phase response. For better illustration, the rangescaled, so that in (e) it is twice and in (f) half the size of (g).


Rese

ARTICLE IN PRESS0.5

1

1.5

8 C. Fermller et al. / Visioncomponent with signicant energy is four times and the temporalcomponent is two times the size of the bar, see Fig. 5). Our analysisis a simulation of the continuous derivation. Clearly, our vision sys-tem does not have motion lters for every position on the retina.But if we assume a non-biased distribution of the lter locations(for example, uniform, or random), we can say that statisticallythe lter outputs should approximate the continuous signal.

5. Experimental evaluation

5.1. Donguri and rotating snakes

The following gures show the estimated image velocity as afunction of spatial frequency. The estimates were obtained by sim-ulations as described in Section 3.3. That is, the motion of a pattern

2 1.5 1 0.5 0 0.5 1 1.5 20

2 1.5 1 0.5 0 0.5 1 1.5 20

0.5

1

1.5

Fig. 9. Velocity estimation at critical frequencies. The dashed (blue) line denotesthe scaled amplitude of ~I. The full (green) line denotes the local estimated velocity(Eq. (9)) and the dot-dashed (red) line denotes the corresponding scaled amplitude.(Note: because the temporal lter estimates the motion at a point earlier in time,the maximum value for I

is found to the right of the stronger edge for left motion

and to the left of the stronger edge for right motion). The estimated average velocity(estimated using Eq. (10)) is larger for left than for right motion. (For interpretationof the references to color in this gure legend, the reader is referred to the webversion of this article.)

Please cite this article in press as: Fermller, C., et al. Illusory motion due to caelement is computed as the energy weighted average of the localvelocities of all the points on the signal (Eq. (10)). Fig. 11a showsthe estimated velocity for left and right motion for a large rangeof frequencies, demonstrating that signicant differences occur ina small range around frequency x 1. Fig. 11b zooms in on aneighborhood around the critical frequencies, but shows only thedifference in velocity between estimated left and right motion.Let us clarify, higher frequencies (of the Gabor lter) in our plotcorrespond to higher resolution of the perceived image, that is l-ters located closer to the fovea.

Referring to Fig. 11a, the estimates uctuate in the neighbor-hood of the critical frequencies. There is an overestimate for leftmotion and an underestimate for right motion at x 1. Bothvelocities are overestimated for a bit larger x (1.25), and bothare underestimated for a bit smallerx (0.75), but at these frequen-cies their differences are not signicant.

To test the validity of the approach, we experimented by vary-ing the parameters in the motion estimation. In particular, we var-ied the range of possible temporal frequencies (with the smallestrange [0.6. . .1.4] and the largest unlimited), the size of the spatialand temporal lters, and the weighting of the local velocity esti-mates. Besides the energy, we used the absolute value of the lterresponse and its cube for weighting. We found that for someparameter settings, both left and right motion were underesti-mated at x 1. However, for all settings, there was a signicantdifference at the critical frequency, with the left motion being lar-ger than the right. Based on these experiments, we state the gist ofour nding as: Estimated left motion is larger than estimated rightmotion for the critical frequencies.

Next, consider the Rotating Snakes pattern (Fig. 2) and takethree cross-sections through one of its units to obtain three quali-tatively different proles. Referring to Fig. 12, the rst cross-sec-tion is at the center, providing a prole just like Donguris, thesecond cross-section is in the upper half, where all intensity re-gions have equal width, and the third is close to the top of the unit,where the intermediate intensity regions (yellow and blue) be-come narrow bars and the black and white regions have large ex-tent. We refer to these proles as Donguri, Bars, and Steps,respectively. Linear arrangements of the corresponding monochro-matic signals are shown in Fig. 12b and c. Fig. 13 plots the differ-ence in estimated image velocity between left and right motionfor the three functions. The simulations show that for all three sig-nals there is a range of frequencies for which left motion is signif-icantly larger than right motion. The difference is signicantlylarger in Donguri than Bars, and is larger in Bars than Steps.

5.2. Nulling experiment

The strength of the illusory perception varies signicantly be-tween observers. The relative strength of the perceived motion indifferent signals, however, can be used to evaluate the model.We quantitatively compared the perception of the three signalsabove by nulling the illusory motion with opposing real motion,similar as in Murakami et al. (2006). Nine naive subjects partici-pated in the experiment.

5.2.1. MethodsThe signals were arranged on three concentric rings, with the

middle ring three times the width of the inner and outer rings.Each ring consisted of 40 signal elements, and the anking ringswere phase shifted with respect to the middle ring by a quarterof the element (Fig. 15). The linearly calibrated intensity valuesof the four regions were 0.3, 1, 0.7, 0, where 0 is black and 1 is

arch xxx (2009) xxxxxxwhite. The width of the bar was 1 unit and the other regions were3 units in Donguri and Stairs, and all regions were 2 units in Bars(as shown in Fig. 12). In addition to these three signals, we also


Rese

ARTICLE IN PRESSC. Fermller et al. / Visiontested a Donguri signal of reduced contrast with the intensities 3/8,1, 5/8, and 0. Simulations for this signal (Fig. 14) show that therange of frequencies with left and right motion being signicantlydifferent is smaller than in the original Donguri, but the predictedvalue at the maximum is nearly the same, actually slightly larger.

Observers were sitting at a distance of 45 centimeters in front ofthe1700 LCDscreenandobserved thepatternsbinocularly.At thisdis-tance the rings covered 14 of visual eldwith thewidth of the threerings covering1.7. At the center of thepatternswas a ring of 1lledwith randomblack andwhite dots for gaze control. Subjectswere in-structed to look inside the disc freely. For each signal two patternswere created, one with the intensity regions in the order shownabove inducing counterclockwise motion, and one obtained as themirror reection of the former, inducing clockwise motion.

Using Matlab, an interface was created that allowed to play vid-eos showing these patterns rotating slowly clockwise or counter-

Fig. 10. Amplitude of Ifor temporal frequencies xt 0:6;1:0 and 1.4. Left motion is cha

is characterized by underestimated temporal frequency on the left edge.

Fig. 11. (a) Velocity estimation for Donguri. (b) Difference in velocity estimates between

Fig. 12. Three different illusory

Please cite this article in press as: Fermller, C., et al. Illusory motion due to caarch xxx (2009) xxxxxx 9clockwise. The speed of motion could be set in the range of0:060:6=s by the step of 0:06, where 1=s corresponds to one de-gree of polar angle per second.

The speed of motion that gave the subjective stationary perceptwas found with the Method of Adjustment. Observers were rstpresented with the static pattern. They then adjusted on a sliderthe speed of motion, upon which a video of the pattern driftingat the selected speed appeared in the location of the static pattern.Observers increased and decreased the speed until they found thespeed, which gave rise to the perception of a stationary pattern.

5.2.2. ResultsFig. 16 displays the measurements. The speed nulling illusory

motion in Donguri was in the range of 0:120:36=s. All participantperceived Donguri the strongest, and Snake stronger than Stairs.None of the subjects measured a signicant difference between

racterized by overestimated temporal frequency on the right edge, and right motion

left and right motion. Higher values of x correspond to higher resolution images.

intensity signals in Snake.


Research xxx (2009) xxxxxx

ARTICLE IN PRESS10 C. Fermller et al. / Visionthe original Donguri and Donguri with decreased contrast, and be-tween clockwise and counterclockwise stimuli. Fig. 17 comparesthe predictions to the mean of measurements. The values shown

are the ratio of the nulling motion in Donguri to Snake, Bars andReduced Donguri, and were obtained as the average over clockwiseand counterclockwise patterns over all subjects. The predicted mo-tions were found as the estimated maximum difference betweenleft and right motion over all frequencies. The gure demonstratesthat our model predicts observed ratios between different condi-tions of the experiment very well.

5.3. Peripheral drift and central drift

It is generally considered that the illusory motion effect wasrst observed in the peripheral drift illusion (Fraser and Wilcox,1979) (see Fig. 18). The intensity prole in this pattern is a saw-tooth function. Such a function would not give rise to erroneouslyestimated motion according to our model. However, luminance re-corded at the neural level is usually modeled as a non-linear func-tion of the actual intensity of the image. Following Backus andOru (2005), we consider two factors in our model: rst, luminance

0.5 1 1.5 2 2.5 3 3.5 40.1

0.05

0

0.05

0.1

0.15

0.2

0.25

Velo

city

diff

eren

ce

Comparison of three patterns

DonguriStepsBars

Fig. 13. Difference in estimated velocity between left and right motion for the threeillusory signals in Snake.

Fig. 14. Difference in estimated velocity for Reduced Donguri.

Fig. 15. Example of experimental stimulus.

Please cite this article in press as: Fermller, C., et al. Illusory motion due to caadaption, a logarithmic function modeling the relationship be-tween recorded luminance and physical intensity (changes at high-er intensity values are recorded with a smaller value than changesat lower intensity values); second contrast adaptation, a sigmoidfunction modeling greater sensitivity to the middle range thanthe high and low ranges of intensities (see Fig. 19).

Kitaoka and Ashida (2004) created a series of patterns, whichthey call central drift illusions, as they are perceived in central aswell as peripheral vision. For examples, see Sakura and Cendri inFig. 21. These patterns contain elements (the petal and ovals) with(close to) linear intensity proles, but in comparison to the periph-eral drift illusion, the individual elements are separated by uniformbackground. This separation increases the illusory effect.

Applying our luminance model to the actual intensities, we ob-tain the luminance proles shown in Fig. 20. Our models predictedvelocity differences between left and right motion for Sakura areshown in Fig. 22. The petals in the model are four units (the barin Donguri is one unit). Thus, the critical frequencies of x 1 inFig. 22 corresponds to the period of the sinusoid being a 14 of the pe-tal size. Fig. 22b compares the velocity differences in peripheraldrift, Sakura, and Cendri for a small range around the critical fre-quencies. Our model predicts perceived illusory motion in the pat-terns, with the effect being stronger in Sakura and Cendri than inperipheral drift.

A similar signal, which Kitaoka (2006) calls Type I, consisting ofeither white to medium gray elements on dark background or

Fig. 16. Cancelation velocity in 9 subjects for clockwise and counterclockwise

illusory motion. Each color corresponds to one subject. (For interpretation of thereferences to color in this gure legend, the reader is referred to the web version ofthis article.)


Rese

ARTICLE IN PRESSBars Steps Red. Donguri0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

Rel

ativ

e ca

ncel

latio

n ve

loci

ty

IIllusory effect relative to Donguri

DataPrediction

C. Fermller et al. / Visionmedium to dark gray elements on white background, causes illu-sory motion in the periphery (Fig. 23a). The corresponding prolesmay be considered smooth versions of the Donguri-prole. A sim-ulation of the motion estimate, considering our luminance modelpredicts the perceived motion (Fig. 23c). In this pattern one ele-ment is chosen three units. Fig. 24 shows one of Kitaokas patternsfrom this class, in which the two elements are combined for aneven stronger effect.

Let us note that contrast and luminance adaptations would noteffect the motion estimation in Donguri and Snake. The estimationis very robust over luminance changes. As shown, the reducedDonguri signal gives rise to very similar motion prediction.

6. 2D motion

The estimation of instantaneous 2D image motion still can beimagined as a two-stage computational process. In the rst stage,causal lters estimate point-wise erroneous motion in the direc-tion perpendicular to the spatial lter orientation (the 1D motioncomponent, also called normal ow, which is the projection of

Fig. 17. Comparison of prediction to experimentally obtained velocity ratios ofBars: Donguri, Steps: Donguri, and Reduced Donguri: Donguri.

Fig. 18. The peripheral drift illusion (Fraser & Wilcox, 1979). In peripheral vision,the circle appears to rotate slowly in clockwise direction.

Please cite this article in press as: Fermller, C., et al. Illusory motion due to caarch xxx (2009) xxxxxx 11the 2D motion vector on the tuning direction of the lter). In thesecond stage normal ow estimates in different directions withinspatial local neighborhoods are combined and the 2D image mo-tion of the patch is estimated.

We implemented the following simple motion algorithm todemonstrate that the residual motion vectors are consistent withthe perceived illusory motion: at the critical frequency of the pat-tern, at every image point we obtain the spatial frequency re-sponses using a standard set of Gabor lters, and we estimatethe corresponding temporal frequency of maximum energy usingcausal lters. Thus, we arrive at n equations of the form

xxi xyiv xti with i 1; . . . ; n: 11

Then we compute the ow u;v of every pattern element by solv-ing the over-determined system of n equations in (Eq. (11)) usingweighted least squares estimation, with the weights the energy re-sponses of the lter outputs.

Fig. 19. Model of neural luminance function.


Rese

ARTICLE IN PRESS12 C. Fermller et al. / VisionFig. 25a and b show for the Donguri pattern the residual owvectors resulting from a horizontal and a vertical movement,respectively. Each of the movements produces image motion onmost of the individual pattern elements. The motion is largest forthe elements with dominant edge direction perpendicular to themovement, but even on elements oriented 60 away from thatdirection there is some image motion. If we combine image mo-tions from only two movements spaced at least 30 apart, we willobtain image motion on all elements. We expect that our visionsystem integrates the motion signals over a time interval of afew eye movements. Since each movement produces image motionin a large range of directions, this process should not be sensitive tothe particular directions of eye movements. Fig. 25c shows the vec-tor sum of the ow elds due to the horizontal and verticalmovements.

Fig. 20. Modeled luminance prole in p

Fig. 21. Sakura in gray: when xating on the center, the outer petals appear tomove slowly clockwise and the inner petals move counterclockwise. B. Cendri(same as Sakura): the ovals on gray background move from gray to white. W.Cendri: the ovals on white background move from white to gray.

Please cite this article in press as: Fermller, C., et al. Illusory motion due to ca7. Discussion

7.1. Signal integration

Estimation of local image motion is the rst step in visual mo-tion analysis. The local signals are input to many visual processes.Some of the very basic processes are the estimation of our ownmo-tion, that is the relative motion of the eye with respect to the scene,and the segmentation of the scene into different objects. Whilecausal lters create local erroneous motion signals in this illusion,the strong perception of rotating patterns is due to these furtherprocesses. First, using the retinal motion signals over the whole vi-sual eld, a 3D motion estimation process obtains the eye move-ments (and the head and body movements if there are any), andstabilizes the image. Second, a segmentation process using as inputthe local motion signals together with information from staticcues, such as edges, texture and color, performs a grouping into cir-cular elements of rotational motion.

According to our motion model all image motion in asymmet-ric signals should be estimated with error. However, the errone-ous estimation in the illusions is due to the motion signal fromdrift movements (Murakami et al., 2006). We speculate that therole of drift movements for this illusion lies in a better temporalintegration of the motion signal when compared to signals fromother movements. We know that images are computationally sta-bilized. The drift motion is computed from the local motion sig-nals over the whole visual eld. Then the drift is discarded, andthe image signals over a time interval are integrated. This is com-putationally feasible, because the drift motion is mostly a rotation

eripheral and central drift patterns.

arch xxx (2009) xxxxxxand does not depend on the structure of the scene. By tting tothe whole image motion eld a rotational motion eld, whichonly depends on three parameters, local motion vectors can beestimated very accurately and reliably. On the other hand, headmotions and scene motions also involve translation, and the im-age motion eld then depends on the scene. Therefore, local mo-tion estimation cannot be that accurate, and integration over atime interval is more difcult.

The illusion appears a bit stronger when viewing the patternsbinocularly versus monocularly. This may be attributed to the re-sponses of binocular motion signals. The drift movements in thetwo eyes are independent. Any single directional movement givesrise to erroneous residual image motion only on some parts of thepatterns (where the edges are perpendicular to the movement, ascan be seen in Fig. 25). Two different drift movements, thus, causeerroneous image motion on more parts, and provide more informa-tion for the integration into rotational motion.

Some observers, and especially many older people do not per-ceive this illusion. We speculate that it is not a lack of involuntaryeyemovements, but decreased sensitivity tomotion in the temporal


0Rese

ARTICLE IN PRESSC. Fermller et al. / Visionhighandmiddle frequency range, ashasbeenmeasured inolderpeo-ple (Shinomori andWerner, 2003; Shinomori and Werner, 2006).

7.2. Relationship to geometric optical illusions

The concept of smoothing at certain scale as an explanation foroptical illusions is not new to the literature. Morgan and Moulden

0

0

0

0

0

Velo

city

diff

eren

ce

Fig. 22. (a) Difference in estimated velocity between left and right motion in Sakura. (b

Fig. 23. (a) Gray shaded elements on bright and dark background. Dark elements on whiopposite to the peripheral drift illusion!) (b) Luminance prole for the dark element on w

Please cite this article in press as: Fermller, C., et al. Illusory motion due to ca.08

0.1Comparison

SakuraW. CendriDrift

arch xxx (2009) xxxxxx 13(1986) and Morgan and Casco (1990) have proposed that bandpassltering (that is edge detection by computing derivatives on asmoothed image) is the cause of a number of (static) geometricoptical illusions. For an example see Fig. 26a. The illusory elementsin this pattern are bars. As discussed in Fermller and Malm, 2004,if we smooth a bar with a Gaussian of r large enough to effect bothedges of the bar but not large enough for the two edges to merge,

0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4.04

.02

0

.02

.04

.06

) Comparison of motion estimation between peripheral and central drift illusions.

te move from dark to light. Light elements on black move from light to dark. (This ishite background. (c) Difference in estimated velocity between left and right motion.


Rese

ARTICLE IN PRESS14 C. Fermller et al. / Visionthe location of the edges changes, as illustrated in Fig. 26c. For abright bar in a dark region (or a dark bar in a bright region) thetwo edges drift apart. For a bar of medium brightness next to abright and a dark region the two edges move toward each other.The latter case corresponds to the Donguri prole. The r in theGaussian is the same as the r in the Gabor of the critical frequen-cies. Thus, at the critical frequencies the interaction of the twoedges causes a change in the location of the edges (dened asthe extrema in the rst-order derivatives or zero-crossings in thesecond-order derivatives). In this paper we showed, that at thesame time local frequencies are poorly estimated, which has an ef-

Fig. 24. Type I illusion

Fig. 25. (a) Estimated residual ow for Donguri at critical frequencies for (a

Fig. 26. (a) Illusory pattern waves a perfect checkerboard pattern with superimposedthe movement of edges under smoothing for a small part of the pattern. (c) A schematic dintensity functions of the two different bars, and the second row shows the proles of tdrift apart or get closer.

Please cite this article in press as: Fermller, C., et al. Illusory motion due to caarch xxx (2009) xxxxxxfect if image sequences are ltered asymmetrically in temporaldomain.

7.3. Summary of the paper

Temporal image motion lters are causal, i.e. they use data fromthe past, but do not use data from the future. Such lters are asym-metric giving greater weight to recent input than older input. Inthis paper we showed that this asymmetry in the lters leads toerroneous estimation of image motion for asymmetric signals atcertain scale. This is simply because of the universal uncertainty

(Kitaoka, 2006).

) horizontal, (b) vertical, (c) combined horizontal and vertical motion.

squares causes the perception of wavy lines (Kitaoka, 1998). (b) Demonstration ofescription of the behavior of edge movement in scale space. The rst row shows thehe (smoothed) functions with the dots denoting the location of edges, which either


in estimating signals. We demonstrated the mis-estimation usingsimulations. Then we tested our model quantitatively using differ-ent signals with bar-like structures and found that it very well pre-dicts the illusory motion perception. Based on these ndings, wehypothesize that this erroneous estimation explains the illusoryperception of motion in static patterns with repeated asymmetricpattern elements under free viewing conditions.

References

Adelson, E. H., & Bergen, J. R. (1985). Spatiotemporal energy models for theperception of motion. Journal of the Optical Society of America A, 2, 284299.

Albrecht, D., & Geisler, W. (1991). Motion selectivity and contrast response functionof simple cells in the visual cortex. Visual Neuroscience, 7, 531546.

Albrecht, D., Geisler, W., Frazor, R., & Crane, A. (2002). Visual cortex neurons ofmonkeys and cats: Temporal dynamics of the contrast response function.Journal of Neurophysiology, 88, 888913.

Anstis, S. (1970). Phi movement as a subtraction process. Vision Research, 10,14111430.

Ashida, H., & Kitaoka, A. (2003). A gradient-based model of the peripheral driftillusion. In Proceedings of the ECVP, Paris.

Backus, B. T., & Oru, I. (2005). Illusory motion from change over time in theresponse to contrast and luminance. Journal of Vision, 5(11), 10551069.

Burr, D., & Morrone, M. (1993). Impulse response functions for chromatic andachromatic stimuli. JOSAA, 10, 1706.

Chen, Y., Wang, Y., & Qian, N. (2001). Modeling V1 disparity tuning to time-varyingstimuli. Journal of Neurophysiology, 86, 143155.

Conway, B., Kitaoka, A., Yazdanbakhsh, A., Pack, C., & Livingstone, M. (2005). Neuralbasis for a powerful static motion illusion. Journal of Neuroscience, 25,56515656.

Eizenman, M., Hallett, P., & Frecker, R. (1985). Power spectra for ocular drift andtremor. Vision Research, 25, 16351640.

Fermller, C., & Malm, H. (2004). Uncertainty in visual processes predictsgeometrical optical illusions. Vision Research, 44, 727749.

Fraser, A., & Wilcox, K. J. (1979). Perception of illusory movement. Nature, 281,565566.

Faubert, J., & Herbert, A. M. (1999). The peripheral drift illusion: A motion illusion inthe visual periphery. Perception, 28, 617621.

Kitaoka, A. (1998). .Kitaoka, A. (2003). .Kitaoka, A. (2006). The effect of color on the optimized FraserWilcox illusion. Gold

prize at the 9th LOR+AL Art and Science of Color Prize.Kitaoka, A., & Ashida, H. (2003). Phenomenal characteristics of the peripheral drift

illusion. Vision, 15, 261262.Kitaoka, A., & Ashida, H. (2004). A new anomalous motion illusion: The central drift

illusion. In Winter Meeting of the Vision Society of Japan.Kuriki, I., Ashida, H., Murakami, I., & Kitaoka, A. (2008). Functional brain imaging of

the rotating snakes illusion by fmri. Journal of Vision, 8(10), 110.Morgan, M. J., & Casco, C. (1990). Spatial ltering and spatial primitives in early

vision: An explanation of the ZllnerJudd class of geometrical illusions.Proceedings of the Royal Society, London B, 242, 110.

Morgan, M. J., & Moulden, B. (1986). The Mnsterberg gure and twisted cords.Vision Research, 26(11), 17931800.

Murakami, I. (2004). Correlations between xation stability and visual motionsensitivity. Vision Research, 44, 251261.

Murakami, I., & Cavanagh, P. (1998). A jitter after-effect reveals motion basedstabilization of vision. Nature, 395, 798801.

Murakami, I., Kitaoka, A., & Ashida, H. (2006). A positive correlation betweenxation instability and the strength of illusory motion in a static display. VisionResearch, 46, 24212431.

Ross, J., Morrone, M., Goldberg, M., & Burr, D. (2001). Changes in visual perception atthe time of saccades. Trends in Neurosciences, 24(2), 113121.

Shi, B. E., Tsang, E. K. C., & Au, P. S. P. (2004). An onoff temporal lter circuit forvisual motion analysis. In ISCAS (Vol. 3, pp. 8588).

Shinomori, K., & Werner, J. (2003). Senescence of the temporal impulse response toa luminous pulse. Vision Research, 43, 617627.

Shinomori, K., & Werner, J. (2006). Impulse response of an S-cone pathway in theaging visual system. Journal of the Optical Society of America A, 23(7), 15701577.

Watson, A. B., & Ahumada, A. J. (1985). Model of human visual motion sensing.Journal of the Optical Society of America, 2, 322342.


ARTICLE IN PRESSPlease cite this article in press as: Fermller, C., et al. Illusory motion due to causal time ltering. Vision Research (2009), doi:10.1016/j.visres.2009.11.021

Illusory motion due to causal time filteringIntroductionMotion estimation in the frequency domainThe filtersModeling the spatial and temporal filtersDefinition of filteringImage motion estimation

The effect of filtering on DonguriExperimental evaluationDonguri and rotating snakesNulling experimentMethodsResults

Peripheral drift and central drift

2D motionDiscussionSignal integrationRelationship to geometric optical illusionsSummary of the paper

References

causal - illusiory motion casual time filtering.pdf

Documents

Transcript of causal - illusiory motion casual time filtering.pdf