Large-Scale Optimization of Hierarchical Features for Saliency Prediction in Natural Images

Transcription

1 Large-Scae Optimization of Hierarchica Features for Saiency Prediction in Natura Images Eeonora Vig Harvard University Michae Dorr Harvard Medica Schoo David Cox Harvard University Abstract Saiency prediction typicay reies on hand-crafted (mutiscae) features that are combined in different ways to form a master saiency map, which encodes oca image conspicuity. Recent improvements to the state of the art on standard benchmarks such as MIT1003 have been achieved mosty by incrementay adding more and more hand-tuned features (such as car or face detectors) to existing modes [18, 4, 22, 34]. In contrast, we here foow an entirey automatic data-driven approach that performs a arge-scae search for optima features. We identify those instances of a richy-parameterized bio-inspired mode famiy (hierarchica neuromorphic networks) that successfuy predict image saiency. Because of the high dimensionaity of this parameter space, we use automated hyperparameter optimization to efficienty guide the search. The optima bend of such mutiayer features combined with a simpe inear cassifier achieves exceent performance on severa image saiency benchmarks. Our modes outperform the state of the art on MIT1003, on which features and cassifiers are earned. Without additiona training, these modes generaize we to two other image saiency data sets, Toronto and NUSEF, despite their different image content. Finay, our agorithm scores best of a the 23 modes evauated to date on the MIT300 saiency chaenge [16], which uses a hidden test set to faciitate an unbiased comparison. 1. Introduction The visua word surrounding us is astonishingy compex. Yet, humans appear to perceive their environment and navigate in it amost effortessy. One bioogica key strategy to reduce the computationa oad and bandwidth requirements is seective, space-variant attention. Effective attentiona mechanisms guide the gaze of the observer to saient and informative ocations in the visua fied. Mimicking such a seective processing has been the subject of in- Now at Xerox Research Centre Europe. Image Human edn Judd CovSa Figure 1. Saiency prediction exampe from the MIT300 chaenge. Our mode, Ensembe of Deep Networks (edn), comes cosest to human data on this benchmark; aso shown are Judd and CovSa (ranked 2nd and 3rd, respectivey). tense research both in neuroscience [15]andincomputervision, where saiency-based preprocessing has found a wide appicabiity from image/video compression and quaity assessment to object and action recognition. Eary agorithms for saiency prediction typicay foowed the Feature Integration Theory [31] andfusedtogether hand-crafted image features such as orientation, contrast, coor extracted on mutipe scaes. These modes differ in their approach to combine individua feature maps into one master map. Many approaches, incudingthecassicaittiandkoch [14] mode, compute normaized center-surround difference maps of the individua features and combine these using a weighting scheme. A more recent fusion scheme expresses feature conspicuity by the equiibrium distribution of a fuy connected graph [10]. Other approaches defined saiency in terms of information theory, e.g. bysef-information[33], information maximization [5], or discriminant saiency that distinguishes target from nu hypotheses [7]. Recenty, spectrum-based methods demonstrated good performance despite ow computationa compexity [12, 11, 28]. Finay, severa data-driven agorithms have been proposed that use ML techniques to predict saiency. Whie Kienze et a. [19] directyearnedacassifierfromfixated image patches, many authors earned the weights associated with a set of predefined features, e.g. [18, 34, 4]. Recent improvements to the state of the art have been achieved mosty by incrementay adding more and more hand-tuned features to such existing modes [18, 4, 22, 34]. For exampe, state-of-the-art performance on the MIT1003 eye-tracking benchmark is achieved by training a cassifier 1

2 on a combination of a arge variety of both ow-eve and high-eve features, incuding object detectors for cars, faces, persons, and the horizon [18, 4]. More recenty, however, image feature earning has gained momentum in the computer vision community, as aresutoftheapproach ssuperiorperformanceonsevera vision tasks ranging from scene cassification to object and face recognition (e.g.[20, 24, 1]). Inthecontext ofsaiency, first attempts empoyed independent component anaysis to earn V1-ike features [5] forgazeprediction,whereasthe authors in [29] earnedhigh-eveconceptsinanunsuper- vised way from fixated image patches. In this work, we expore the usefuness of such bioinspired hierarchica features to predict where peope ook in natura images. Our approach is structured as foows. We generate a arge number of instances of a richyparameterized bio-inspired hierarchica mode famiy, and seect those that are predictive of image saiency. The combination of severa independent modes shoud improve performance [27], but a brute-force search for the best mixture is computationay prohibitive. Therefore, we use hyperparameter optimization [2, 3] tospeedupthesearch both for individua modes and their combinations and find a discriminative image representation for saiency. We then demonstrate that a simpe inear cassifier that operates on such representation outperforms the state of the art on the MIT1003 benchmark [18]. Moreover, these representations that were earned on MIT1003 generaize we to two other image saiency data sets, Toronto [5] and NUSEF [26], despite their different image content. Additionay, we show that our mode outperforms a 22 agorithms evauated to date on the MIT300 saiency benchmark (see Figure 1 and [16]). Our resuts demonstrate that aricher,automaticay-derivedbaseofhierarchicafeatures can chaenge the state of the art in saiency prediction. In summary, we make severa contributions to the fied. First, we introduce hierarchica feature earning to the area of visua saiency, yieding state-of-the-art resuts on four benchmarks, incuding an independent third-party evauation with a hidden test set. Furthermore, we impemented an automated approach to optimize hierarchica features for saiency prediction, as opposed to a more traditiona use of hand-tuned features. Finay, we make pubicy avaiabe the software to compute our saiency maps at 2. Bio-inspired saiency features We start by reviewing the broad cass of bioogicayinspired visua representations that we use here and present adjustments to this architecture to suit the task at hand. We then outine ways to efficienty search this vast representation space for instances that are particuary discriminative for saiency. We combine these buiding bocks and describe our feature earning pipeine in Section Richy-parameterized mutiayer visua representations Typica saiency features are hand-tuned and have a oose connection to the architecture of bioogica visua systems. To derive more compex, bioogicay more pausibe saiency features, we consider a broad cass of bio-inspired hierarchica modes [24] thathaspreviousybeenshownto exce in various recognition tasks from face verification [24] and identification to object recognition [3]. These modes beong to the more genera cass of convoutiona neura networks [21] and,accordingy,haveahierarchicamuti- ayer structure. Inspired by the organizationa principes in the primate visua cortex, hierarchies are assembed from inear fitering and non-inear transformation stages. More specificay, each network ayer is comprised of asetofoperationsthatcorrespondtobasicmechanisms known to take pace in the visua cortex. These operations are parameterized by a arge number of architectura variabes, hence, are highy configurabe. We can here ony give a very brief overview and refer the reader to [24, 25]for more detais of exact ayer ayout, range of parameters, and ademonstrationoftheeffectivenessofsuchrepresentations in the context of face recognition. The set of operations: 1. Linear fitering by convoution with a bank of random uniform fiters: Fi = N 1 Φ i, where N 1 is the normaized input (mutichanne image or feature map) of ayer, {1, 2, 3}, andφ i,i {1,...,k },is arandomfiter. Thisoperationproducesastackofk feature maps F. Parameters: fiter shapes s f s f,s f {3, 5, 7, 9} and fiter count k {16, 32, 64, 128, 256] 2. Activation with a bounded activation function: A = Activate(F ) such that Activate(x) = γmax γmin x if x>γmax if x<γmin otherwise Parameters: γmin {, 0}, γ max {1, + } 3. Spatia smoothing by pooing over a spatia neighborhood a a : P = Poo(A ) such that ( ) P p = Downsampe α (A ) p 1 a a (2) Parameters: a {3, 5, 7, 9}, exponent p {1, 2, 10}, α downsamping factor 4. Loca normaization by the activity of neighbors across space (b b neighborhood) and feature maps k : N = Normaize(P ) such that N = { C Ĉ 2 ρ C if ρ Ĉ 2 > τ otherwise (1) (3)

3 where C = P δ ˆP with δ {0, 1} controing whether or not the oca mean ˆP is subtracted, and Ĉ = C 1 b b k Parameters: ρ stretching param., τ threshod, δ, b Any individua instantiation of this famiy of visua representations can contain an arbitrary number of ayers that are stacked on top of each other. To constrain the computationa compexity, we here consider ony one- to three-ayer (L1 L3) feature extractors (aso caed modes ). In the foowing, we describe our changes to generaize this architecture to the task of saiency prediction. A major difference of this work is reated to the genera purpose of these features. The origina approach aimed at obtaining a compact, fixed-size representation of an image that can then be fed into cassifiers for a singe, goba decision (e.g. faceidentity).incontrast,hereweseekocaized representations of images. We keep the input image at high resoution and abe individua pixes of the output feature map by their saiency. Furthermore, the modes in [24] were imited to grayscae input. Because of the importance of coor in determining the saiency of a region we extend the modes to mutispectra input. Additionay, we aso consider the YUV coor space, which provides decorreated uminance and chrominance channes (simiar to the coor space empoyed by the human visua system). YUV has been shown to give better resuts than RGB for saiency prediction (e.g. [28]). In addition to evauating individua modes, we augment our feature set by representation bending. Bycombining together mutipe good representations one can take advantage of the structura diversity of the individua modes in the bend. Such bending strategies have been shown to significanty boost performance [27, 24] Guided search for optima saiency features The guiding principe in the architectura design of the above mode famiy was configurabiity. To aow for a great variety of feed-forward architectures, the mode cass is richy parameterized. Depending on the number of ayers, modes have up to 43 architectura parameters (see Sec. 2.1). The performance of any singe mode instantiation may range from chance to state-of-the-art performance depending on parameter configurations. To find good representations for a specific task, we perform a arge-scae search over the space of a possibe mode candidates, which are cross-vaidated on a separate screening set. In principe, modes can be drawn entirey randomy. Indeed, random search was found to be an effective approach to search for particuary discriminative representations for recognition tasks [24]. However, for arge enough probems, random search is sti prohibitivey computationay expensive. For exampe, in [24], good representations were found ony after an exhaustive search of amost 13,000 modes. Therefore, the mode search shoud not be random but guided towards better regions of the parameter space. Recenty, automated hyperparameter optimization [2, 3] was proposed to use Bayesian optimization methods to guide search in a arge parameter space. In an object recognition scenario, these optimization agorithms achieved the same resuts as an exhaustive random search agorithm, in a fraction of the time required by random search [3]. Here, we use the pubicy avaiabe toobox of Bergstra et a. [3] to more efficienty search the vast bioinspired mode space for optima saiency features. This method invoves defining (i) a search space (as an expression graph) and (ii) a oss function to be minimized. In addition to a description of the bio-inspired mode cass, the search space aso contains the hyperparameters of cassifiers, such as the strength of reguarization. The oss function we use is defined in the next section Feature earning pipeine To evauate the performance of our bioogicay-inspired representations, we foow the standard saiency earning pipeine of Judd et a. [18]. This offers a standardized way to evauate new features and makes our approach directy comparabe with the baseine method of [18]. Saiency prediction is formaized as a supervised earning probem. A training set of saient (i.e.thepositivecass) and non-saient (negative cass) image regions is obtained from ground-truth empirica saiency maps (aso caed fixation density maps) derived from rea eye movement data. For each image in the training set, we randomy pick 10 saient sampes from the top 20% saient regions and 10 non-saient sampes from the bottom 70% saient areas of the empirica saiency map. At these seected ocations, features are extracted from the image and normaized (over the entire training set) to zero mean and unit variance. Finay, the abeed feature vectors are fed into an L2-reguarized, inear, L2-oss SVM, which is trained to predict for each ocation in a new test image its probabiity of fixation (see Fig. 2). To search for good representations for saiency prediction, we consider a subset of 600 images of the MIT1003 eye movement data set [18], and perform mode search on this set. To estimate the prediction error, we perform 6- fod cross-vaidation. In the testing phase, we consider the continuous output w T x + b of the SVM. By threshoding these continuous saiency maps, image regions above the threshod are cassified as saient. A systematic variation of the threshod eads to an ROC curve; the oss function to be minimized by the hyperparameter optimization agorithm (Sec. 2.2)isthen1-AUC(i.e.theAreaUndertheROC Curve). Note that both the architectura and earning parameters (parameter C of the inear SVM) were tuned simuta-

4 Test images Fixation map L3 Feature extraction TESTING TRAINING Optima bend of L1, L2, L3 features: Train images Normaize Labeed feature vectors L1 Normaize 1 Linear SVM... L2 Normaize edn saiency map Figure 2. Schematic diagram of our pipeine. Good Li mutiayer feature extractors are found by guided hyperparameter search (not shown) and combined into an optima bend. Resuting feature vectors are abeed with empirica gaze data and fed into a inear SVM. For detais on operations inside each i ayer see Sec neousy by hyperparameter optimization. To identify those instances (or combinations thereof) of the above described bio-inspired mode famiy that successfuy predict image saiency, we adopt a two-step search procedure. First, we search for we-performing individua L1, L2, and L3 modes, keep these and perform another search for optima combinations of these seected modes. Our fina saiency mode is an ensembe of individua modes, hence the name Ensembe of Deep Networks (edn). In the first screening stage, we independenty screened approximatey 2000 L1, 2200 L2, and 2800 L3 modes on RGB input. For the sake of comparison, a fraction of these modes was screened with random search ( 1200 L2 and 2400 L3 modes) and the rest with the more efficient guided search (Sec. 2.2). We partiay rey on random search, because the best configurations returned by this type of search tend to be ess correated than those found by guided search. The rationae is that a greater diversity in individua modes may have a positive effect on the performance of mode bends. In addition, we separatey screened about 1900 L1, 1700 L2, and 1900 L3 modes on YUV input. Of these, about 600 L3 modes were drawn randomy. In the second search stage, we seected 5 to 10 topperforming modes from each mode cass (a cass being defined by the number of ayers, the input coor space, and the type of screening), and screened unconstrained combinations of up to 8 of these modes. The best ensembe was found after around 1000 trias and consists of 6 architecturay diverse modes: three L3-RGB modes found with random search, and one L2-YUV and two L3-YUV modes found with guided search. We note that other combinations, incuding those with L1 modes, were aso found among top-performing resuts. Fig. 3 shows exampe outputs of individua modes and the optima bend. The entire screening was performed on a arge CPU custer and took approximatey 1 week. Figure 3. Output of best individua L1, L2, L3 (YUV) modes and the optima bend. Left to right, top: input, histogram-matched edn saiency map (bend of 6 Li modes), human saiency map; Midde: raw L1, L2, and L3 output; Bottom: L1, L2, L3 output histogram-matched to the human saiency map. Whie L1 modes detect edges, L2 and L3 capture higher-order structures Center bias and smoothness Severa studies have pointed out that gaze is biased to the image center (often irrespective of the center features). More and more modes expicity incorporate this bias (e.g. [17, 32]). We foow [18] and extend our earned features with a simpe 1-D measure of the distance of each pixe to the image center. Because some reference modes do not account for this bias, for a fair comparison, we incorporate the same distance-to-center feature in these modes as we. In addition, we report resuts aso without this center feature. However, it shoud be noted that some agorithms aready have an impicit center bias due to border effects. As a fina step, saiency maps are often smoothed to bet-

5 ter match the distribution of eye movements [17]. We determine the optima Gaussian bur eve (σ=30px) through cross-vaidation on a subset of the training set. 3. Evauation: eye movement prediction Because of the tight ink between saiency, attention, and eye movements, saiency modes are typicay evauated in terms of how we they predict human gaze in natura scenes. We foow the same approach here Data sets Feature search and performance evauation were conducted on the MIT1003 [18] dataset. Totestthegenera- izabiity of the mode, we use three additiona benchmarks, Toronto [5], NUSEF [26], and MIT300 [16]. The MIT1003 data set [18] consistsof1003imagesand the eye movement data of 15 observers who free-viewed these images. Because the feature search stage used 600 of these images for training, the remaining 403 served as test images. The Toronto benchmark [5] containseyemovement data from 20 subjects watching 120 images. It is considered a hard data set due to the ack of particuary saient regions (such as faces/peope). The recenty proposed NUSEF [26] ismadeupof758semanticay-rich images incuding affective content (expressive faces, nudes, unpeasant concepts). A tota of 75 observers viewed parts of the image set. We here ony consider the pubicy avaiabe part (441 images). Finay, for an independent evauation, we submitted our mode for third-party evauation on the MIT300 saiency benchmark [16]. This set consists of 300 images viewed by 39 observers whose gaze data is not pubic. To date, 23 saiency modes have been evauated on this patform Evauation protoco We foow the evauation procedure outined in Sec In addition to ROC anaysis, for the sake of comprehensiveness, we consider three other common evauation metrics: the Earth Mover s Distance (EMD, in the context of saiency see [34]), Normaized Scanpath Saiency (NSS) [23], and a simiarity score [17]. As reference for comparisons, we consider 11 stateof-the-art saiency agorithms: GBVS [10], the mutiscae quaternion DCT signature on YUV input (denoted QDCT) [28], Judd [18], AWS [8], CovSa [6] (withco- variances + means), Tavakoi [30], AIM [5], Goferman [9], ICL [13], Image Signature (with LAB coor space) [11], and Boost [4], a with defaut parameters. This seection is based on the top-performing modes from two recent and comprehensive reviews/taxonomies [4, 16]. To estimate the effective performance range, we used two contro measures. First, we computed eave-one-sub- Mode RGB YUV L L L edn Tabe 1. AUC scores of best individua modes and the optima ensembe of top modes on MIT1003. Performance increases with mode compexity (i.e. # ayers) and the use of YUV. The optima ensembe of Deep Networks (aso found through automated screening) gives highest performance. No center bias at this point. ject-out empirica saiency maps and used these to predict the eye movements of the eft-out viewer. Since peope are sti the best predictors of where other peope ook, this measure constitutes an upper bound for prediction. Our ower bound comes from the above-mentioned center bias (e.g. [32]): a simpe measure based on the distance of each pixe to the image center predicts eye movements we Resuts on the MIT1003 benchmark First, we anayzed individua mode performance by systematicay varying two meta-parameters: the number of ayers in each mode (1, 2, or 3) and the input coor space (RGB or YUV). Performance (AUC) of the best cassspecific modes is shown in Tabe 1. Consistent with expectations, performance increases with mode compexity (L1 < L2 < L3), i.e. modes with more ayers achieve higher invariance. Confirming previous findings, use of the YUV coor space gives significanty higher prediction performance than RGB. For the choice of the search agorithm, guided search not ony is more efficient than random search (360 iterations give better resuts than 2400 w/o guidance) but aso exceeds the best random search performance. The most significant performance boost is achieved with ensembes of mutipe good representations (see Tabe 1, edn). In contrast to [25], however, bends are not imited to the combination of the top modes, but derived again in an automated fashion through guided screening. Bending is so beneficia because combinations take advantage of the diversity of individua modes. This expains why individuay weaker modes (such as L3-RGB or L2-YUV) are aso represented in the best bend. One aspect of diversity is the scae on which individua modes operate. From the combination of mutipe modes tuned to different scaes, a mutiscae mode emerges. Performance of the various saiency agorithms (baseines, contros, and our best bend) is summarized as averages over a MIT1003 test images in the top part of Tabe 2. edn outperforms a individua saiency agorithms for a four metrics. A sma further performance gain (for AUC and EMD) can ony be achieved by Boosting [4], i.e. optimay bending many top-performing agorithms from this

6 MIT1003 Toronto Nusef AUC EMD sim NSS w/ C w/o C w/ C w/o C w/ C w/o C w/ C w/o C Chance Cntr QDCT ICL CovSa Signature GBVS Tavakoi Goferman AIM Judd AWS edn Boost Boost+eDN Human Chance Cntr CovSa ICL Signature GBVS QDCT Goferman Tavakoi AIM Judd AWS edn Boost Boost+eDN Human Chance QDCT Cntr Tavakoi ICL GBVS Signature CovSa Goferman Judd AIM AWS edn Human Tabe 2. Performance of saiency agorithms with (w/ C) and without (w/o C) center bias on the MIT1003, Toronto, and NUSEF benchmarks for four metrics: AUC, simiarity (sim), Normaized Scanpath Saiency (NSS) (the higher the better for a three) and EMD (ower better). Feature search ony used AUC as objective function. edn outperforms individua modes on MIT1003, on which features and cassifier are earned. A sma further improvement can ony be achieved by bending mutipe individua top-performing agorithms (see Boost [4] and Boost+eDN). edn generaizes we to Toronto and NUSEF, despite their different image content. Sma border artifacts (due to repeated fitering) make edn inherenty biased to the center hence the performance advantage in w/o C case.

7 tabe (Judd [18], GBVS [10], AWS [8] and Torraba). We aso note that Boosting resuts get better with incusion of edn features, e.g AUCforMIT1003(seeTabe2). Many image processing steps used in saiency computations introduce border artifacts (see [33]foradetaiedanaysis of severa agorithms). Because of repeated fitering across mutipe ayers, our saiency maps suffer from some border effects as we. Hence, simiary to GBVS, even our non-centered maps ( w/o C in Tabe 2) are sti impicity biased to the center and have a performance advantage. Exampe saiency maps for our agorithm and some reference methods are shown in Fig Generaization to other data sets To assess how we our earned features generaize to other data sets, we first evauated them on the Toronto and NUSEF benchmarks. Resuts are shown in Tabe 2 (ower part). Despite being trained on a different data set, edn outperforms the state-of-the-art approaches. This is surprising, considering the significant database-specific differences in image content. NUSEF was deiberatey created to investigate semantics-driven attention to affective content. Conversey, Toronto acks particuar regions of interest (peope, faces), so that gaze is ess coherent across viewers. Finay, on the MIT300 saiency benchmark, our mode achieves AUC ( simiarity and EMD), sighty better than the second best mode of Judd et a. with AUC (0.506 simiarity and 3.13 EMD) see [16]. 4. Discussion and concusion Hierarchica feature earning has become a common approach in computer vision, but has not been adequatey expored in the context of saiency prediction. Here, we addressed this issue and efficienty searched a arge poo of richy parameterized neuromorphic modes for those representations that are discriminative for saiency. Through automated representation bending, we assembed powerfu combinations of diverse mutiayer architectures that outperform the state of the art. In notabe contrast to topperforming hand-tuned modes, our approach makes no assumptions about what ower-eve features (coor, contrast, etc.) or higher-eve concepts (faces, cars, text, horizon) attract the eyes. Instead, we aow the hierarchica modes to earn such compex patterns from gaze-abeed natura images (e.g. seetoprowinfigure4). We beieve our integrated and bioogicay-pausibe approach is therefore conceptuay ceaner and more generic than approaches that rey on domain-specific hand-crafted features. Athough trained ony on part of the MIT1003 benchmark, our representations generaize we to three other eye movement data sets, in spite of their different image content. Despite the arge size of the mode space (i.e. the mode famiy is as incusive as possibe), good candidates are found quicky (within a coupe of hundred trias) through nove hyperparameter optimization agorithms. Our resuts show that methods empoying rich, automaticay-derived feedforward representations can chaenge the state of the art in the fied of saiency prediction. Acknowedgments. This work was supported by the Nationa Science Foundation (IIS ) and a gift from Inte Corporation. Eeonora Vig was supported by a feowship within the Postdoc-Programme of the German Academic Exchange Service (DAAD, D/11/41189). References [1] Y. Bengio, A. Courvie, and P. Vincent. Representation earning: A review and new perspectives. arxiv preprint arxiv: , [2] J. Bergstra, R. Bardenet, Y. Bengio, and B. Kég. Agorithms for hyper-parameter optimization. In NIPS 25, , 3 [3] J. Bergstra, D. Yamins, and D. Cox. Making a science of mode search: Hyperparameter optimization in hundreds of dimensions for vision architectures. In ICML,2013.2, 3 [4] A. Borji. Boosting bottom-up and top-down visua features for saiency estimation. In CVPR,2012.1, 2, 5, 6 [5] N. Bruce and J. Tsotsos. Saiency based on information maximization. In NIPS 18,pages , 2, 5 [6] E. Erdem and A. Erdem. Visua saiency estimation by nonineary integrating features using region covariances. Journa of vision,13(4), [7] D. Gao and N. Vasconceos. Bottom-up saiency is a discriminant process. In ICCV,pages1 6, [8] A. Garcia-Diaz, X. R. Fdez-Vida, X. M. Pardo, and R. Dosi. Decorreation and distinctiveness provide with human-ike saiency. In Advanced Concepts for Inteigent Vision Systems,pages ,2009.5, 7 [9] S. Goferman, L. Zenik-Manor, and A. Ta. Context-aware saiency detection. PAMI, 34(10): , [10] J. Hare, C. Koch, and P. Perona. Graph-based visua saiency. In NIPS 19,pages ,2007.1, 5, 7 [11] X. Hou, J. Hare, and C. Koch. Image signature: Highighting sparse saient regions. PAMI,34(1),2012.1, 5 [12] X. Hou and L. Zhang. Saiency detection: A spectra residua approach. In CVPR,pages1 8, [13] X. Hou and L. Zhang. Dynamic visua attention: Searching for coding ength increments. In NIPS, [14] L. Itti, C. Koch, and E. Niebur. A mode of saiencybased visua attention for rapid scene anaysis. IEEE PAMI, 20(11): , [15] L. Itti, G. Rees, and J. K. Tsotsos, editors. Neurobioogy of Attention. Esevier, San Diego, CA, [16] T. Judd, F. Durand, and A. Torraba. MIT saiency benchmark. 1, 2, 5, 7

8 Image Human edn Judd GBVS Boost AWS AIM Figure 4. Sampe images from MIT1003 [18] with ground truth (Human), proposed (edn) and reference saiency maps. Saiency maps were histogram-matched to the corresponding fixation density map for visuaization. edn maps are more consistent with human maps. [17] T. Judd, F. Durand, and A. Torraba. A benchmark of computationa modes of saiency to predict human fixations. MIT tech report,2012.4, 5 [18] T. Judd, K. Ehinger, F. Durand, and A. Torraba. Learning to predict where humans ook. In ICCV, pages , , 2, 3, 4, 5, 7, 8 [19] W. Kienze, F. A. Wichmann, B. Schökopf, and M. O. Franz. A Nonparametric Approach to Bottom-Up Visua Saiency. In NIPS,pages , [20] A. Krizhevsky, I. Sutskever, and G. Hinton. Imagenet cassification with deep convoutiona neura networks. In NIPS 25, pages , [21] Y. LeCun and Y. Bengio. Convoutiona networks for images, speech, and time series. The handbook of brain theory and neura networks,3361, [22] Y. Lu, W. Zhang, C. Jin, and S. Xue. Learning attention map from images. In CVPR,pages , [23] R. J. Peters, A. Iyer, L. Itti, and C. Koch. Components of bottom-up gaze aocation in natura images. Vision research, 45(18): , [24] N. Pinto and D. D. Cox. Beyond Simpe Features: A Large-Scae Feature Search Approach to Unconstrained Face Recognition. IEEE Automated Face and Gesture Recognition, , 3 [25] N. Pinto, D. Doukhan, J. J. DiCaro, and D. D. Cox. A highthroughput screening approach to discovering good forms of bioogicay inspired visua representation. PLoS Comput Bio,2009.2, 5 [26] S. Ramanathan, H. Katti, N. Sebe, M. Kankanhai, and T.- S. Chua. An eye fixation database for saiency detection in images. In ECCV,pages30 43,2010.2, 5 [27] R. E. Schapire. The boosting approach to machine earning: An overview. Noninear Estimation and Cassification. Lecture Notes in Statist.,pages ,2003.2, 3 [28] B. Schauerte and R. Stiefehagen. Quaternion-based spectra saiency detection for eye fixation prediction. In ECCV, pages , , 3, 5 [29] C. Shen, M. Song, and Q. Zhao. Learning high-eve concepts by training a deep network on eye fixations. In NIPS Deep Learning and Unsup Feat Learn Workshop, [30] H. R. Tavakoi, E. Rahtu, and J. Heikkiä. Fast and efficient saiency detection using sparse samping and kerne density estimation. In Image Anaysis,pages [31] A. M. Treisman and G. Geade. A feature-integration theory of attention. Cognitive psychoogy, 12(1):97 136, [32] P. Tseng, R. Carmi, I. Cameron, D. Munoz, and L. Itti. Quantifying center bias of observers in free viewing of dynamic natura scenes. Journa of Vision,9(7),2009.4, 5 [33] L. Zhang, M. H. Tong, T. K. Marks, H. Shan, and G. W. Cottre. SUN: A Bayesian framework for saiency using natura statistics. Journa of Vision, 8(7):1 20, , 7 [34] Q. Zhao and C. Koch. Learning a saiency map using fixated ocations in natura scenes. Journa of Vision, 11(3), , 5