Analyzing Intrinsic Motion Textures
`Created from Naturalistic Video Captures
`
`Angus Graeme Forbes1, Christopher Jette1 and Andrew Predoehl2
`1School of Information: Science, Technology, and Arts, University of Arizona, Tucson, United States
`2Department of Computer Science, University of Arizona, Tucson, United States
`angus.forbes@sista.arizona.edu, cnj@email.arizona.edu, predoehl@email.arizona.edu
`
`Keywords:
`
`Intrinsic Motion Textures, Psychophysics, Perception, Metrics for Motion, Dynamic Textures.
`
`Abstract:
`
`This paper presents an initial exploration of the plausibility of incorporating subtle motions as a useful modal-
`ity for encoding (or augmenting the encoding of) data for information visualization tasks. Psychophysics
`research indicates that the human visual system is highly responsive to identifying and differentiating even the
`subtlest motions intrinsic to an object. We examine aspects of this intrinsic motion, whereby an object stays in
`one place while a texture applied to that object changes in subtle but perceptible ways. We hypothesize that the
`use of subtle intrinsic motions (as opposed to more obvious extrinsic motion) will avoid the clutter and visual
`fatigue that often discourages visualization designers from incorporating motion. Using transformed video
`captures of naturalistic motions gathered from the world, we conduct a preliminary user study that attempts
`ascertains the minimum amount of motion that is easily perceptible to a viewer. We introduce metrics which
`allow us to categorize these motions in terms of flicker (local amplitude and frequency), flutter (global ampli-
`tude and frequency), and average maximum contrast between a pixel and its immediate neighbors. Using these
`metrics (and a few others), we identify plausible ranges of motion that might be appropriate for visualization
`tasks, either on their own or in conjunction with other modalities (such as color or shape), without increasing
`visual fatigue. Based on an analysis of these initial preliminary results, we propose that the use of what we
`term “intrinsic motion textures” may be a promising modality appropriate for a range of visualization tasks.
`
`1
`
`INTRODUCTION
`
`Rustling leaves, flickering flames, sunlight sparkling
`on water – every day we are continually confronted
`with naturalistic motion as we navigate the world.
`Information visualization, as a field, examines how
`meaning is effectively conveyed through visually-
`encoded data. However, dynamic visualizations that
`encode data using motion have not been as widely ex-
`plored, as motion is generally considered to be too
`distracting a modality for representing information
`effectively. The human visual system is exception-
`ally adept at identifying differences in texture (Em-
`rith et al., 2010), and observing both the movement
`of objects or the movement within an object (Chalupa
`et al., 2004). While far from being fully understood,
`a growing body of research indicates the presence of
`multiple neurological mechanisms for processing ex-
`trinsic motion (the movement of an object) and intrin-
`sic motion (local movement within a single object)
`(Lu and Sperling, 2001; Nishida et al., 1997). Extrin-
`sic and intrinsic motion is also termed “first-order”
`
`or “second-order,” respectively: first-order motion re-
`ferring to the perception of a change in luminance
`across the visual field; second-order motion referring
`instead to the perception of changes in texture or con-
`trast. While both types of motion can indicate a global
`movement of a single entity, in this paper we examine
`the use of second-order motion to indicate motions
`within a stationary object. We introduce a method for
`transforming real-world motion into abstract motion
`textures with subtle motions, which we are calling in-
`trinsic motion textures. These transformed video cap-
`tures of real-world naturalistic motion allow us to ex-
`periment with non-distracting motion without includ-
`ing their representational aspects. Through the trans-
`formation of, say, the movement of water in a stream,
`we are able to capture the motion without directly re-
`minding users that they are looking at water. We also
`introduce an easy-to-calculate set of metrics to char-
`acterize these intrinsic motions. While the vast range
`of possible motions makes it rather daunting to at-
`tempt to encompass all types of movement via a sin-
`gle set of metrics, ours capture the main features of
`
`Lightricks Ltd.
`EX1018
`Page 1 of 7
`
`

`

`Figure 1: Stages of the video capture of naturalistic motion being transformed into an intrinsic motion texture. Here we show
`a single frame (A) as it is (B) cropped, (C) desaturated, (D) contrast-compressed with a low pixel range, and (E) pixelated. A
`further step uses temporal smoothing to mitigate extra flickering introduced from the pixelation.
`
`intrinsic motion. We conducted a user study asking
`participants to evaluate a set of intrinsic motion tex-
`tures. We were primarily interested in gathering in-
`sight into the following question – What is the least
`amount of motion needed in order to easily identify
`differences between highly similar motions? We dis-
`covered that even intrinsic motion textures with a low
`contrast range are easily distinguishable, as long as
`certain amounts of flicker, flutter, and average maxi-
`mum contrast (defined below) are present.
`In the influential text, “Semiology of Graphics,”
`Jacques Bertin introduces a typology of retinal vari-
`ables applicable to the communication of informa-
`tion. Although explicitly discussing static print me-
`dia, Bertin hints at the ability of texture, unlike other
`retinal variables, to produce a vibratory effect that is
`the “collusion” of physiological and psychological ef-
`fects. Although he does not explore this issue of vi-
`bratory effects further, he encourages designers “to
`make the most of this variation, to obtain the res-
`onance without provoking an uncomfortable sensa-
`tion, to flirt with ambiguity without succumbing to
`it.” Bertin speaks of texture in terms of the vary-
`ing thicknesses of lines, points, or shapes to indicate
`an overview similarity or differentiation between dif-
`ferent qualities, or an ordered-ness within the same
`quality. However, he does not believe that the use of
`texture is refined enough to allow viewers to perceive
`proportionality and therefore is not effective at allow-
`ing users to perform quantitative tasks using texture
`alone (Bertin, 2010).
`Nonetheless, investigations into the use of static
`textures in visualization contexts have found that they
`can be effective for representing quantitative data in
`certain tasks. A seminal paper (Ware and Knight,
`1995) uses Gabor patches to parametrically investi-
`gate aspects of texture that may be useful in informa-
`tion display, such as orientation, size, and contrast.
`Applied research introduced in (Kujala and Lukka,
`2003) explores the use of a “parameter hierarchy” in
`the procedural creation of perceptually distinct tex-
`tures for the effective display of information. More
`
`recently, (Byelas and Telea, 2009) uses overlapping
`textures to indicate multivariate data in software dia-
`grams, explicitly attempting to indicate an increased
`number of variables within the diagram. (House et al.,
`2006) describes a technique for the perceptual opti-
`mization of complex visualizations involving layered
`textures. And (Interrante and Kim, 2001) and (Kim
`et al., 2004) explore the efficacy of using various ori-
`entations and types of textures to facilitate shape per-
`ception. These examples are mostly explicitly con-
`cerned with the potential expanded range of variables
`that can be expressed, while at the same time aware
`that the cost of this expanded range might be percep-
`tual fatigue or cognitive overload, or worse, an in-
`ability to clearly distinguish differences. (Interrante,
`2000) examines the use of overlapping naturalistic
`textures to indicate multivariate data while mitigating
`against the “extraneous stress” that might occur with
`synthesized textures. In addition to being a potentially
`effective representation of a wide range of quantita-
`tive information, the use of certain aspects of natu-
`ralistic textures, such as variability, might be used to
`indicate extra information, such as uncertainty. Inter-
`estingly, recent perceptual experiments, such as (Em-
`rith et al., 2010), confirm that humans perceive even
`minute alterations in texture, but note that it is in fact
`somewhat easier to discern differences between syn-
`thetic textures than natural textures.
`Motion is often used to signal transitions between
`views and contexts, to signal interruptions, and to in-
`dicate temporal aspects of data. However, it is less
`frequently used as an encoding mechanism for quanti-
`tative or qualitative information. (Forbes et al., 2010)
`presents a data visualization framework than enables
`animation to be mapped to dynamic streams of data,
`and (Bostock et al., 2011) describes a framework
`that includes “transition” operators for animating data
`points. The use of motion in visualization elicits con-
`cern about visual clutter and perceptual fatigue, even
`while potentially providing an expanded toolset for
`representing information. At its most extreme, the in-
`judicious use of motion in information might cause
`
`Page 2 of 7
`
`

`

`significant visual stress (Ware, 2004). One group
`of security researchers (Conti et al., 2005) even de-
`scribes the potential for malicious hackers to takeover
`an information visualization system and alter its vi-
`sual output to induce epileptic seizures.
`Results from (Bartram and Ware, 2002) show that
`small, brief, and graphically simple extrinsic motions
`are perceptually efficient ways to distinguish objects
`in a crowded display. In particular, they note that a
`synchronization of elements is required in order for
`them to be effectively recognized as similar. That is,
`the timing of the motion is as important as the mo-
`tion itself. Research into extrinsic motion cues, or
`“moticons” (Bartram et al., 2003), finds that motion
`coding is independent from color and shape coding
`and that more subtle motions are less distracting to
`users yet easily perceived. A series of experiments
`that analyzed extrinsic aspects of motion– velocity,
`direction, and on-off blinking– finds that these prop-
`erties are all effective at encoding multiple data values
`in a prototype astrophysics simulation, provided they
`meet certain basic thresholds of perceptibility (Huber
`and Healey, 2005). A technique termed “motion high-
`lighting” explores the potential applicability of mo-
`tion to node-link diagrams (Ware and Bobrow, 2004).
`Results of motion highlighting experiments indicate
`that the translating or scaling of node is more useful
`for supporting rapid interactive queries on node-link
`diagrams than static highlighting methods.
`Dynamic textures are sequences of images that
`exhibit some form of temporal coherence, or more
`specifically, they are individual images that are “re-
`alizations of the output of a dynamical system driven
`by an independent and identically distributed pro-
`cess” (Doretto et al., 2003). Dynamic textures have
`been effectively used in scientific visualizations and
`have been more extensively investigated in computer
`graphics and computer vision research. For instance,
`(Van Wijk, 2002) uses an iterative series of texture
`distortions to represent fluid flows, and (Forbes and
`Odai, 2012; Forbes et al., 2013) applies this tech-
`nique to creative media arts projects. Work by (Lum
`et al., 2003) explores adding moving particles to a
`surface texture of a static object in which the parti-
`cles are placed along the principal curvature direction
`to better indicate the object’s shape and spatial rela-
`tionships. Within a more general computer graphics
`context, dynamic textures are used in a variety of ap-
`plications. For instance, dynamic textures have been
`used as a computationally efficient way to add realism
`to a scene. (Chuang et al., 2005) presents an interac-
`tive system that procedurally generates dynamic tex-
`tures from selected components of a single image that
`can then be added to a scene. Similarly, a variety of
`
`techniques have been introduced to automatically cre-
`ate “temporal textures” from a single image in order
`to mimic natural phenomenon such as clouds, water,
`and fire (Lai and Wu, 2007; Ruiters et al., 2010; Ok-
`abe et al., 2011). In addition to having the potential
`to be used as an effective modality for representing
`quantitative information, recent research has explored
`the use of dynamic textures as a medium for provid-
`ing semantically contextualized information. (Lock-
`yer et al., 2011) explores the “expressive scope” of
`ambient motion textures for “emphasis and more sub-
`tle ambient visualization.” In particular, this research
`focused on the effective communication of particu-
`lar emotions through the use of intrinsic motion cues
`within a dynamic texture.
`For the most part, research on motion in informa-
`tion visualization is concerned with extrinsic motion,
`or at least does not differentiate between extrinsic and
`intrinsic motion. For instance, (Ware and Bobrow,
`2004), also cited above, discusses a motion highlight-
`ing technique whereby the animation of a station-
`ary link generate a “crawling” motion. Although it
`is not presented specifically as a dynamic texture, it
`is clear that this “crawling” motion is of a different
`nature than the translation patterns used to highlight
`nodes. A recent evaluation found that animated rep-
`resentations were more effective than almost all static
`representations of link representations (Holten et al.,
`2011).
`It seems reasonable that other visualization
`systems could utilize a conflation of textures and mo-
`tion; rather than attempting to procedurally generate
`dynamic textures, we could gather them directly. This
`would have the immediate advantage that they were,
`at least in some degree, inherently non-distracting for
`the simple reason that they occur continually in the
`real-world. An earlier (unpublished) study by one of
`the authors found that the use of moving sinusoidal
`gratings introduced visual fatigue precisely because
`of the qualities that made it unrealistic: its fixed ro-
`tation, its predictable frequency and amplitude, and
`its repetitive sequence of pixel values.
`instead, Dy-
`namic textures using fluctuating, intrinsic, real-world
`motion are cohesive without being repetitive; differ-
`entiable without being distracting.
`In order to expedite the creation of intrinsic mo-
`tion textures in order to analyze their potential effec-
`tiveness in visualization systems, we gathered real-
`world video of natural phenomena containing intrin-
`sic motion: fire, water, clouds, etc. Although we be-
`lieve that these textures have a representational com-
`ponent that might be useful in some visualization cir-
`cumstances, for this study we isolated the phenomena
`using these steps: (A) record the natural phenomenon;
`(B) crop the resulting video; (C) desaturate the video;
`
`Page 3 of 7
`
`

`

`The variability metrics reflect the spatial variation
`in a single video frame. Let Q′ (cid:26) Q be the set of
`pixel positions in the interior of the grid. Each pixel
`position q 2 Q′ therefore has eight spatially adjacent
`neighbors, the set of which we denote N(q). The
`roughness R is defined as the average intensity dif-
`ference of the highest-contrast neighbors:
`Definition 2 (Roughness).
`T(cid:229)
`(cid:1) 1
`(cid:229)
`T
`q2Q′
`Edginess Eq is the average number of large-
`contrast juxtapositions, per pixel, per frame. The con-
`trast is regarded as large if the intensity difference is
`at least q:
`Definition 3 (Edginess).
`T(cid:229)
`(cid:1) 1
`1
`(cid:229)
`W H
`T
`q2Q′
`t=1
`In the current study, we set the threshold value q equal
`to one-eighth the maximum contrast range of the rele-
`vant collection (G20, G40, G60, or G80). For example,
`when comparing two videos from group G40, we used
`the E40=8 = E5 edginess metric.
`Our flicker metrics depend on the local maxima
`and minima (peaks and valleys) of pixel intensity in
`the time domain. To specify them formally, we intro-
`duce definitions of peak and valley as follows.
`Definition 4 (Peak intensity value). Iq;t is a peak in-
`tensity value of width j, provided
`(cid:15) j > 0,
`(cid:15) 1 < t (cid:20) T (cid:0) j,
`(cid:15) Iq;t(cid:0)1 < Iq;t = Iq;t+1 = (cid:1)(cid:1)(cid:1) = Iq;t+ j(cid:0)1, and
`(cid:15) Iq;t+ j(cid:0)1 > Iq;t+ j.
`this definition is consistent
`Though cumbersome,
`with an intuitive notion of local maximum. We also
`define valley intensity value analogously for local
`minima. Let nq denote the number of peaks of width
`1 or more at position q. Let p1;q; p2;q; : : : ; pnq;q be the
`peak intensity values at position q, such that pi;q is the
`intensity value of the ith peak, in chronological order.
`Similarly, let v1;q;v2;q; : : : ;vnq;q denote the valley in-
`tensity values (assuming they are equally numerous
`as the peaks). Now we can precisely state our met-
`rics for flicker. First, the average number of peaks per
`pixel, per frame, is called local frequency, FL:
`Definition 5 (Local frequency).
`(cid:1) 1
`1
`W H
`T
`
`nq:
`
`(cid:229) q
`
`2Q
`
`FL =
`
`R =
`
`1
`W H
`
`jIq;t (cid:0) Is;tj:
`
`max
`s2N(q)
`
`t=1
`
`Eq =
`
`1ft : q(cid:20)maxs2N(q)
`
`jIq;t(cid:0)Is;tjg(t):
`
`(D) constrain the pixel range of the video; (E) pixelate
`the video; (F) apply temporal smoothing to the video.
`The video is thus transformed into a low-contrast, de-
`saturated, pixelated, and mostly unrecognizable ver-
`sion of itself that nonetheless retains important qual-
`ities of the original natural motion. Figure 1 shows
`a frame from a naturalistic video as it is processed
`through this pipeline. These intrinsic motion textures
`can then be defined metrically, in terms of particular
`features, and included in studies where we can asso-
`ciate these features with empirical observations, such
`as discernibility and differentiability.
`
`2 Metric definitions
`
`Much research has been done to develop ways to ef-
`fectively and efficiently characterize motion. A com-
`monly used method, optical flow, assumes that there
`is a unique velocity vector at each pixel. Other meth-
`ods relax that assumption. For instance, (Langer and
`Mann, 2003) introduces “optical snow,” which is able
`to characterize motions that have prevalent disconti-
`nuities between frames. The intrinsic motions that
`we have gathered likewise include large amounts of
`flickering that are not captured with optical flow type
`analyses. Since we are, for now, interested primar-
`ily in the single task of determining the ease of dis-
`crimination betweens motions, we constructed a way
`to create simpler metrics that define a video via a set
`of eight features that sufficiently characterize intrinsic
`motions. They are grouped into the following cate-
`gories: flicker, flutter, and variability.
`We now introduce notation that defines our met-
`rics precisely. Let Iq;t denote the integer intensity
`value of a pixel in the video, at spatial position q
`and at time t. Assuming the video pixels lie in a
`rectangular grid of width W , height H, and that the
`video has duration of T discrete video frames, q 2
`Q = f1; : : : ;Wg(cid:2)f1; : : : ;Hg and t 2 f1; : : : ;Tg. The
`first characteristic of interest we define is the contrast
`{
`{
`}
`range K of a video:
`Definition 1 (Contrast range).
`K = max
`Iq;t
`q2Q
`q2Q
`1(cid:20)t(cid:20)T
`1(cid:20)t(cid:20)T
`In the present study, instead of using K as a video
`feature for comparison, we partitioned our test videos
`into collections with of similar contrast range, be-
`cause two videos with widely differing contrasts are
`very obviously distinct. Specifically, we gathered
`videos with contrast ranges of 20 or less, 21 to 40,
`41 to 60, and 61 to 80, into groups called G20, G40,
`G60, and G80, respectively.
`
`Iq;t
`
`:
`
`}(cid:0) min
`
`Page 4 of 7
`
`

`

`Again, in our study we set the this metric’s threshold
`just as those of Defs. 3 and 7. So, when comparing
`videos from group G40, we compute metric CG;5.
`These metrics are easy to calculate and capture
`both the global and the local, pixel level aspects of
`the videos. They may not however capture larger-
`scale movement within the motion textures. However,
`since our express aim is to use motion textures that do
`not contain large-scale motion, we believe that these
`metrics are appropriate as a first attempt to character-
`ize intrinsic motions.
`
`3 User Study
`
`The main goal of our preliminary user study was
`to determine the minimum amount of movement re-
`quired in order for a participant to quickly differen-
`tiate between similar motions. Since motion can be
`highly distracting and since humans are exception-
`ally good at noticing differences in motion, by finding
`lower bounds on various parameters that make motion
`distinguishable we can identify the minimum values
`of easily-discernible features. Future work will use a
`more rigorously defined empirical study using tech-
`niques to measure just-noticeable difference, as well
`as explore user response to visualization tasks incor-
`porating motion textures. For this preliminary study
`we wanted to obtain an initial sense of what attributes
`were most easily noticeable at low-contrast ranges,
`and which of these attributes were thought to be the
`least distracting.
`To find this minimum feature set, we created a
`study that presented the participant with a pair of
`videos. The user was then asked to indicate whether
`he or she agreed or disagreed with a series of state-
`ments about the videos. We gathered 32 unique
`videos of naturalistic motion and processed them as
`described in section 3.1. We created 4 “bins” and
`made versions of each of these videos with different
`levels of contrast. Bin 1 contained videos with a con-
`trast range of +/- 10; Bin 2, +/- 20; Bin 3, +/- 30; Bin
`4, +/- 40. For Bin 1, it was very difficult to tell most of
`the videos apart, especially when looking at a single
`(unmoving) frame from the video. In other words, the
`contrast was so low that without movement it would
`be almost impossible to tell them apart. For Bin 2,
`it seemed that about half of the time it was easy to
`tell the videos apart and the other half of the time it
`was difficult. For Bin 3, it became much easier to tell
`any of the videos apart from any of the others. And
`finally, for Bin 4 it was easy to tell all of the videos
`apart. However, we thought that if the movements be-
`came more chaotic (higher absolute flicker amplitude
`
`We define the average peak-to-valley difference,
`per pixel, as local amplitude, AL:
`Definition 6 (Local amplitude).
`nq(cid:229)
`(pi;q (cid:0) vi;q);
`
`i=1
`
`1 n
`(cid:229) q
`
`q
`
`1
`W H
`
`AL =
`
`2Q
`if nq > 0, otherwise AL = 0.
`We define local choppiness, CL;q, as the average
`number of large intensity jumps per pixel, per frame.
`A jump at position q is large if equals or exceeds
`threshold value q.
`Definition 7 (Local choppiness).
`T(cid:229)
`(cid:1) 1
`(cid:229)
`T
`q2Q
`Similar to our use of Def. 3, in our study we set
`the threshold value here to one-eighth of the contrast
`range of whichever videos we are comparing. So
`when comparing videos from group G40, we use lo-
`cal choppiness metric CL;5.
`Our flutter metrics are similar to the flicker met-
`rics, but they depend on the average intensity of the
`entire video frame at a given moment. Let Jt denote
`the average pixel intensity of the frame at time t, i.e.,
`W H (cid:229)q2Q Iq;t. This sequence also has peaks and
`Jt = 1
`valleys, which we assume are m in number. At the
`risk of confusion, we will denote its peak values and
`valley values (regardless of width) as p1; p2; : : : ; pm
`and v1;v2; : : : ;vm respectively, in chronological order.
`(Note that these values have only one subscript.) The
`average number of these peaks, per frame, is called
`global frequency, FG:
`Definition 8 (Global frequency).
`
`CL;q =
`
`1
`W H
`
`1ft : jIq;t(cid:0)Iq;t(cid:0)1j(cid:21)qg(t)
`
`t=2
`
`FG = m=T:
`
`The average of these peak-to-valley differences is
`the global amplitude, AG:
`Definition 9 (Global amplitude).
`m(cid:229)
`(pi (cid:0) vi);
`
`AG =
`
`1 m
`
`i=1
`if m > 0, otherwise AG = 0.
`We define global choppiness CG;q as the average
`number of large increases in average intensity, with
`threshold q:
`Definition 10 (Global choppiness).
`T(cid:229)
`
`1ft : jJt(cid:0)Jt(cid:0)1j(cid:21)qg(t):
`
`t=2
`
`1 T
`
`CG;q =
`
`Page 5 of 7
`
`

`

`and frequency) then in those cases the videos in Bin
`3 and Bin 4 would be hard to tell apart. We did not
`test any of the videos against a video with a different
`range of pixel values as it is easy to discern the dif-
`ferences in videos when one had a higher maximum
`and lower minimum pixel value than the other. Each
`of the videos was analyzed with custom software that
`output the features described by our metrics system.
`We further calculated the absolute difference between
`the feature vectors of each video.
`We included a series of four Likert items per test
`designed to elicit the participant’s opinion about the
`discernibility of flicker and frequency. We ran vari-
`ous “batches” of our test over the course of a week
`and a half on Amazon Mechanical Turk. We received
`a total of 144 completed studies. For most of these
`batches, we randomly chose one of the 4 bins for each
`test. The majority of our “workers,” 107, indicated
`that they were from India; 24 were from the United
`States; the rest came from the United Kingdom, Mex-
`ico, Sri Lanka, Canada, Pakistan, and Nigeria. 2 par-
`ticipants chose “Other” as their nationality. There
`were an equal number of male and female participants
`(72 each). The minimum and maximum age was 19
`and 63, respectively, with a median age of 31. Fol-
`lowing the suggestions in (Heer and Bostock, 2010),
`which describes some of the advantages and disad-
`vantages of conducting studies via Mechanical Turk,
`we made a substantial effort to encourage reliable par-
`ticipation and mitigate inaccurate or random answers,
`ultimately obtaining 476 samples from the 144 partic-
`ipants.
`All features, except for the frequency of the flutter
`(the global frequency of a direction change in aver-
`age pixel value for a frame) were positively correlated
`with easy differentiability. We built a statistical model
`to characterize the relationship between the video mo-
`tion metrics and the participants’ responses. We fo-
`cused on contrast ranges 20 and 40 and modeled the
`data as a two-category classification problem: in each
`video comparison, the videos are either difficult to
`distinguish (category CD) or not (category CE ), gen-
`erated by the user’s Likert responses. For our binary
`classifier, any value greater than 2 was given a placed
`in category CE, and any value less than or equal to 2
`was placed in category CD. Ideally, the model would
`effectively predict the category for a video compar-
`ison, based only on our video features. We modeled
`these two categories by assuming a multivariate Gaus-
`sian distribution of the feature vectors (which are the
`absolute differences of each of the eight metrics for
`the pair of videos being compared). In other words,
`we computed the maximum-likelihood mean vector
`µC;r and covariance matrix SC;r of all feature vec-
`
`tors for category C 2 fCD;CEg, and contrast range
`r 2 f20;40g. For the purpose of classification, we also
`make use of the empirical frequency of CD and CE
`classes, denoted p(CD) and p(CE ). Given a new data
`vector x for contrast range r, we would classify it in
`category CD provided it satisfies p(CDjx) > p(CEjx),
`(
`)
`where by Bayes’ theorem, for C 2 fCD;CEg,
`(
`)
`x;µC;r;SC;r
`N
`p(C)
`p(Cjx) =
`x;µB;r;SB;r
`(cid:229)B2fCD;CEg N
`p(B)
`In the above, N (x;µ;S) denotes the probability den-
`sity function for the multivariate Gaussian with mean
`µ and covariance S. One advantage of a Gaussian
`characterization is that we can easily marginalize any
`subset of features. Thus we can see the average in-
`teraction between any two features and can list the
`thresholds for the classifier with all other features
`marginalized (Table 1). In particular, even small dif-
`ferences in flickering (especially in the frequency and
`choppiness) at the individual pixel level were the main
`predictors of whether or not a video pair was likely to
`be easily differentiable.
`
`:
`
`Table 1: Threshold values between features.
`l chop
`0.0011901
`l amp
`0.12419
`l freq
`0.00081854
`g chop
`0.048523
`g amp
`0.037790
`g freq
`0.0056176
`rough
`0.57185
`edge
`0.0033416
`
`4 Conclusion
`
`This paper presents an initial foray into exploring
`the potential usefulness of intrinsic motion textures.
`We provide a method for generating cohesive, non-
`repetitive, intrinsic motions textures from real-world
`video captures; a method for characterizing the fea-
`tures of intrinsic motions; a preliminary user study
`that indicates minimal differences necessary for dif-
`ferentiation between motions; and an analysis of this
`study that identifies thresholds on these features. This
`initial exploration of motion textures created from
`video captures of naturalistic movement seems to in-
`dicate that this may be a promising area for future in-
`vestigations. Future work will involve the design and
`analysis of more rigorous empirical studies to deter-
`mine the validity of our claims regarding the notice-
`ability and distraction of these types of textures.
`
`Page 6 of 7
`
`

`

`REFERENCES
`
`Bartram, L. and Ware, C. (2002). Filtering and brushing
`with motion. Information Visualization, 1(1):66–79.
`Bartram, L., Ware, C., and Calvert, T. (2003). Moticons:
`detection, distraction and task. International Journal
`of Human-Computer Studies, 58(5):515–545.
`Bertin, J. (2010). Semiology of graphics: diagrams, net-
`works, maps, pages 79–81. Esri Press.
`Bostock, M., Ogievetsky, V., and Heer, J. (2011). D3: Data-
`IEEE Transactions on Visualiza-
`driven documents.
`tion and Computer Graphics (TVCG), 17(12):2301–
`2309.
`Byelas, H. and Telea, A. (2009). Visualizing multivariate
`attributes on software diagrams. In Software Mainte-
`nance and Reengineering, 2009. CSMR ’09. 13th Eu-
`ropean Conference on, pages 335 –338.
`Chalupa, L., Werner, J., and of Technology, M. I. (2004).
`The visual neurosciences. MIT Press.
`Chuang, Y., Goldman, D., Zheng, K., Curless, B., Salesin,
`D., and Szeliski, R. (2005). Animating pictures with
`stochastic motion textures. ACM Transactions on
`Graphics (TOG), 24(3):853–860.
`Conti, G., Ahamad, M., and Stasko, J. (2005). Attacking in-
`formation visualization system usability overloading
`and deceiving the human. In Proceedings of the 2005
`symposium on Usable privacy and security, pages 89–
`100. ACM.
`Doretto, G., Chiuso, A., Wu, Y. N., and Soatto, S. (2003).
`Dynamic textures. International Journal of Computer
`Vision, 51:91–109. 10.1023/A:1021669406132.
`Emrith, K., Chantler, M., Green, P., Maloney, L., and
`Clarke, A. (2010). Measuring perceived differences in
`surface texture due to changes in higher order statis-
`tics. JOSA A, 27(5):1232–1244.
`Forbes, A. G., H¨ollerer, T., and Legrady, G. (2010). Be-
`haviorism: A framework for dynamic data visualiza-
`IEEE Transactions on Visualization and Com-
`tion.
`puter Graphics (TVCG), 16(6):1164–1171.
`Forbes, A. G., H¨ollerer, T., and Legrady, G. (2013). Genera-
`tive fluid profiles for interactive media arts projects. In
`Proceedings of the International Symposium on Com-
`putational Aesthetics in Graphics, Visualization, and
`Imaging (CAe), pages 123–129, Anaheim, California.
`Forbes, A. G. and Odai, K. (2012). Iterative synaesthetic
`In Proceed-
`composing with multimedia signals.
`ings of the International Computer Music Conference
`(ICMC), pages 573–578, Ljubjiana, Slovenia.
`Heer, J. and Bostock, M. (2010). Crowdsourcing graphical
`perception: using mechanical turk to assess visualiza-
`In ACM Human Factors in Computing
`tion design.
`Systems (CHI), pages 203–212. ACM.
`Holten, D., Isenberg, P., van Wijk, J., and Fekete, J. (2011).
`An extended evaluation of the readability of tapered,
`animated, and textured directed-edge representations
`in node-link graphs. In Pacific Visualization Sympo-
`sium (PacificVis), 2011 IEEE, pages 195 –202.
`House, D., Bair, A., and Ware, C. (2006). An approach
`to the perceptual optimization of complex visualiza-
`
`tions. IEEE Transactions on Visualization and Com-
`puter Graphics (TVCG), 12(4):509–521.
`Huber, D. and Healey, C. (2005). Visualizing data with
`motion. In Proceedings of IEEE Visualization (VIS),
`pages 527–534.
`Interrante, V. (2000). Harnessing natural textures for multi-
`variate visualization. Computer Graphics and Appli-
`cations, IEEE, 20(6):6 –11.
`Interrante, V. and Kim, S. (2001). Investigating the effect
`of texture orientation on the perception of 3d shape.
`In Human Vision and Electronic Imaging VI, volume
`4299, pages 330–339.
`Kim, S., Hagh-Shenas, H., and Interrante, V. (2004). Con-
`veying shape with texture: Experimental investiga-
`tions of texture’s effects on shape categorization judg-
`ments. IEEE Transactions on Visualization and Com-
`puter Graphics (TVCG), 10(4):471–483.
`Kujala, J. and Lukka, T. (2003). Rendering recognizably
`unique textures. In In