Learning Non-Linear Reconstruction Models for Image Set Classification

Size: px
Start display at page:

Download "Learning Non-Linear Reconstruction Models for Image Set Classification"

Transcription

1 Testing Training Learning Non-Linear Reonstrution Moels for Image Set Classifiation Munawar Hayat, Mohamme Bennamoun, Senian An Shool of Computer Siene an Software Enginnering The University of Western Australia {mohamme.bennamoun, Abstrat Training image sets Class Speifi Moels We propose a eep learning framework for image set lassifiation with appliation to fae reognition. An Aaptive Deep Network Template (ADNT is efine whose parameters are initialize by performing unsupervise pretraining in a layer-wise fashion using Gaussian Restrite Boltzmann Mahines (GRBMs. The pre-initialize ADNT is then separately traine for images of eah lass an lass-speifi moels are learnt. Base on the minimum reonstrution error from the learnt lass-speifi moels, a maority voting strategy is use for lassifiation. The propose framework is extensively evaluate for the task of image set lassifiation base fae reognition on Hona/UCSD, CMU Mobo, YouTube Celebrities an a Kinet ataset. Our experimental results an omparisons with existing state-of-the-art methos show that the propose metho onsistently ahieves the best performane on all these atasets.. Introution Fae reognition has traitionally been onsiere as a single image lassifiation problem. With the reent avanes in imaging tehnology, multiple images of a person are beoming reaily available in numerous senarios suh as vieo base surveillane, multi-view amera networks, personal albums an images of a person aquire over a long perio of time. Fae reognition from these multiple images is formulate as an image set lassifiation problem an has gaine a signifiant attention from the researh ommunity in reent years [7, 7, 5, 4, 8, 4, 6]. Compare with single image base lassifiation methos, fae reognition from image sets offers more promises as it an effetively hanle a wie range of variations that are ommonly present in the faial images of a person. These variations inlue hanging illumination onitions, view point variations, expression eformations, olusions an isguise. Faial images of a person uner ifferent variations are ommonly moele on a non-linear mani- Image Set Test image set Image Set k Reonstrution For eah image Voting Θ( Θ(k Output: Class Label ytest Figure. The blok iagram of the propose metho. During training, lass-speifi moels are learne from images of eah person. These moels are then use by a reonstrution error base voting strategy to eie about the lass of a test image set. fol geometry suh as Grassmannian manifol [7, 5, 8] or Lie Group of Riemannian manifol [6]. This moeling of images on manifols requires prior assumptions relate to the speifi ategory of the manifol on whih fae images are believe to lie. In ontrast, this paper introues a eep learning base framework whih makes no prior assumption regaring the unerlying geometry of fae images an an automatially learn an isover the struture of the omplex non-linear surfae on whih fae images of a person (uner ifferent variations are present. The propose framework first efines an Aaptive Deep Network Template (ADNT whose weights are initialize by unsupervise layer-wise pre-training using Gaussian Restrite Boltzmann Mahines (GRBMs. The pre-initialize ADNT is then separately traine for images of eah lass to learn lass-speifi moels. The training is performe in a way that the ADNT learns to reonstrut images of that lass. A lass-speifi moel is therefore mae to learn the struture an the geometry of omplex non-linear surfae on whih fae images of that lass are present. For lassifiation, a reonstrution error an maority voting base strategy is evise. The propose framework is evaluate for vieo base fae reognition on Hona/UCSD [8], CMU Mobo [6] an YouTube Celebrities atasets [6] as well as a Kinet ataset [9, 5] an ahieves state of the art performane.

2 . Relate Work Image set lassifiation generally involves two maor steps:. to fin a representation of the images in the set, an. to efine suitable istane metris for the omputation of the similarity between these representations. Base on the use type of representation, existing image set lassifiation methos an be ategorize into parametri-moel an non-parametri-moel methos. The parametri-moel methos [] approximate an image set in terms of the parameters of a ertain statistial istribution moel an then measure the similarity between two image sets (two istribution parameters using e.g. KL-ivergene. These methos fail to proue a esirable performane if there is no strong statistial relationship between the test an the training image sets. The other type of image set representation methos i.e. non-parametri methos o not make any assumption about the statistial istribution of the ata. These methos have shown promising results an are being atively evelope reently. The non-parametri moel base methos represent an image set either by its representative exemplars or on a geometri surfae. Base upon the type of representation, ifferent istane metris have been evelope to etermine the between-set istane. For example, for the image sets represente in terms of representative exemplars, the set-set istane an be efine as the Euliean istane between the set representatives. These an simply be the set mean [7] or aaptively learnt set samples [4]. Cevikalp et al. [4] learn the set samples from the affine hull or onvex hull moels of the set images. The set to set istane is then terme as Affine Hull Image Set Distane (AHISD or Convex Hull Image Set Distane (CHISD. Hu et al. [4] efine set-set istane as the istane between their Sparse Approximate Nearest Points (SANPs. SANPs of two sets are etermine from the mean image an the affine hull moel of the orresponing set an are sparse approximate from the set images while simultaneously searhing for the losest points in the respetive sets. As set representative base methos require the omputation of a one-to-one set istane, these methos are apable of hanling intra set variations very effetively. However, their performane is highly prone to outliers. They are also omputationally very expensive as a one-to-one math of the query set with all sets in the galley is require. These methos oul therefore be very slow in the ase of a large gallery size. Unlike set representative base methos, the seon ategory of non-parametri methos moel a omplete image set as a point on a geometri surfae [7, 5, 8, 6, 0, 3]. The image set an be represente either by a subspae, mixture of subspaes or on a omplex non-linear manifol. Prinipal angles have been very ommonly use to etermine the istane between image sets represente by a linear subspae. The prinipal angles 0 θ θ π between two subspaes are efine as the smallest angles between any vetor in one subspae an any other vetor in the seon subspae. The similarity between sets is then efine as the sum of the osines of the prinipal angles. For image set representations on manifols, appropriate istane metris have been aopte suh as the geoesi istane [] an the proetion kernel metri [7] on the Grassmann manifol, an the log-map istane metri [9] on the Lie group of Riemannian manifol. In orer to isriminate image sets on the manifol surfae, ifferent learning strategies have been evelope. Mostly, Linear Disriminant Analysis (LDA is ontrive for ifferent set representations. Examples inlue Disriminative Canonial Correlations (DCC [7], Manifol Disriminant Analysis (MDA [5], Graph Embeing Disriminant Analysis (GEDA [8] an Covariane Disriminative Learning (CDL [6]. The methos whih moel an image set on a geometri surfae make prior assumption about the unerlying surfae on whih the fae ata lies. For example, [7] assumes that fae images lie on a linear surfae an represents the image set as a linear subspae. Methos inluing MMD, MDA an GEDA represent an image set on a non-linear Grassmannian manifol, whereas, CDL [6] represents an image set in terms of the ovariane matrix of pixel values on Lie Group of Riemannian manifol. For our propose metho, we o not make any prior assumptions about the struture of the surfae on whih the faial images of a person lie. We instea efine a eep learning base framework whih inorporates non-linear ativation funtions to automatially learn the unerlying manifol struture. Deep learning has reently gaine signifiant researh attention in a number of areas [, 3, 5]. Ours is the first metho whih inorporates eep learning for image set lassifiation. The etaile esription about our metho is presente next. 3. Propose Tehnique We first efine an Aaptive Deep Network Template (ADNT whih will be use to learn the unerlying struture of the ata. The arhiteture of our ADNT is summarize in Fig an the etails are presente in Se 3.. For suh a eep network to perform well, an appropriate initialization of the weights is require. We initialize the weights of the ADNT by performing pre-training in a greey layer wise fashion using Gaussian Restrite Boltzmann Mahines (etails in Se 3.. The ADNT with preinitialize weights is then separately fine-tune for eah of the k lasses of the training image sets. We therefore en up with a total of k fine-tune eep network moels, eah orresponing to one of the k lasses. The fine-tune moels are then use for image set lassifiation (etails in Se 3.3

3 W e ( 04 W e ( Enoer 400 W e (3 00 W ( Deoer 400 W ( 04 W (3 Figure. Struture of the Aaptive Deep Network Template (ADNT. The parameters of the template are initialize by unsupervise pre-training. The initialize template is then use to learn lass speifi moels 3.. The Aaptive Deep Network Template (ADNT As epite in Fig, our ADNT is an Auto-Enoer (AE, onsisting of two parts: an enoer an a eoer. Both the enoer an the eoer have three hien layers eah, with a share thir layer (the entral hien layer. The enoer part of the AE fins a ompat low imensional meaningful representation of the input ata. We an formulate the enoer as a ombination of non-linear funtions s(. use to map the input ata x to a representation h given by, h = s(w e (3 h + b (3 e h = s(w e ( h + b ( e h = s(w e ( x + b ( e Where W e (i R i i is the enoer weight matrix for layer i having i noes, b (i e R i is the bias vetor an s(. is the non-linear ativation funtion (typially a sigmoi or tangent hyperboli. The enoer parameters are learnt by ombining the enoer with the eoer an ointly training the enoer-eoer struture to reonstrut the input ata by minimization of a ost funtion. The eoer an therefore be efine as a ombination of nonlinear funtions whih reonstrut the input x from the enoer output h. The reonstrute output x of the eoer is given by, x = s(w (3 x + b (3 x = s(w ( x + b ( x = s(w ( h + b( ( ( We an represent the omplete enoer-eoer struture (the ADNT by its parameters θ ADNT = {θ W, θ b }, { } 3 { } 3. where θ W = W e (i, W (i an θb = b e (i, b (i Later (in Se. 3.3 we will use this template an separately train it for all lasses of the training image sets to learn lass speifi moels. 3.. ADNT s Parameter Initialization The above efine ADNT is use to learn lass speifi moels. This is aomplishe by separate training of the ADNT for images of eah lass of the training image sets. The training is performe with stohasti graient esent through bak propagation. The training fails if the ADNT is initialize with inappropriate weights. More speifially, if the initialize weights are too large, the network gets stuk in loal minima. On the other han, if the initialize weights are too small, the vanishing graient problem is enountere uring bak propagation in the initial layers an the network beomes infeasible to train. The weights of the template are therefore initialize by performing unsupervise pre-training []. For that, a greey layer-wise approah is aopte an Gaussian RBMs are use. Below, we first present a brief overview of binary an Gaussian RBMs an then explain their use for our ADNT s parameter initialization. An RBM is a generative unirete graphial moel with a bipartite struture of two sets of binary stohasti noes terme as the visible ({v i } Nv, v i {0, } an the hien layer noes ({h } N h, h {0, }. The noes of the visible layer are symmetrially onnete with the noes of the hien layer through a weight matrix W R Nv N h but there are no intra layer noe onnetions. The oint probability p(v, h of the RBM struture is given by, p(v, h = exp( E(v, h (3 Z Z is the partition funtion (use as a normalization onstant an E(v, h is the energy funtion of the moel efine as: E(v, h = i b i v i h i w i v i h (4 Where b an are the biases of the visible an hien layer noes respetively. The training of an RBM for learning its moel parameters {W, b, } is performe by Contrastive Divergene (CD, a numerial metho propose by Hinton et al. [, 3] for effiient approximation to graient omputation an RBM parameter learning. The stanar RBM evelope for binary stohasti ata an be generalize to the real value ata by appropriate moifiations in its energy funtion. Guassian RBM

4 (GRBM is one suh popular extension whose energy funtion is efine by moifying the bias term of the visible units as: E GRBM (v, h = (v i b i σ h v i w i h i i σ i i (5 σ i is the stanar eviation of the real value Gaussian istribute inputs to the visible noe v i. It is possible to learn σ i for eah visible unit but this beomes iffiult when using CD for GRBM parameter learning. We instea aopt an alternative approah an fix σ i to a unit value in the ata pre-proessing stage. Due to the restrition that there are no intra-layer noe onnetions, inferene beomes reaily tratable for the RBM as oppose to most irete graphial moels. The probability istributions for GRBM are given by, where p(h = v = sigmoi ( i w iv i + p(v i h = σ i π exp( (vi ui σi (6 u i = b i + σi w i h (7 Sine our ata is real value, we use GRBMs to initialize the weights of our ADNT. Two layers are onsiere at a time an the GRBM parameters are learnt. Initially, the noes of the input layer are onsiere to be visible units v an the noes of the first hien layer as the hien units h of the first GRBM an its parameters are learnt. The ativations of the first GRBM s hien units are then use as an input to train the seon GRBM. The proess is repeate for all three hien layers of the enoer part of the ADNT struture. The weights learnt for the enoer layers are then tie to the orresponing eoer layers i.e. W (3 = W e ( T (, W = W e ( T (, W = W e (3 T (See Fig. for notations 3.3. Image Set Classifiation Algorithm We are now reay to esribe our reonstrution error base image set lassifiation algorithm. The omplete algorithm is summarize in Alg. The etails are presente below. Problem Formulation: Given k training image sets {X } k an their orresponing lass labels y [,, k], where the image set X = { x (t} has N N images x (t R x y belonging to lass, the problem of image set lassifiation an be formulate as follows: given a test image set X test = { x (t} N test, fin the lass y test to whih X test belongs to? Unsupervise Pre-Training: We first efine our ADNT an initialize its weights by performing unsupervise pretraining. Our ADNT is a multi-layer neural network with noes. In orer to initialize the weights of the ADNT by GRBMs, we generate an unsupervise training ata set. Fae images from all training image sets are gathere into a ata set X u = { x (t X ; [,, k] }. The images in the resulting ata set X u are ranomly shuffle an use for layerwise GRBM training of all layers of the enoer part of the template ( The weights of the eoer layers ( are then initialize with their orresponing tie weights of the enoer layers. Using pretraining for weights initialization has several avantages over ranom initialization. Sine the ADNT is pre-traine for fae images, the initialize weights are very lose to the atual weights []. Therefore, it is highly unlikely that the network gets stuk in a loal optima. Moreover, with properly initialize weights, the graient omputation beomes feasible resulting in the onvergene of the weights to optimal values. Learning Class Speifi Moels: Now that we have the ADNT struture with pre-initialize weights, we separately fine tune its parameters θ ADNT = {θ W, θ b } for eah of the k training image sets. We therefore learn k lass-speifi moels. The learning of a lass-speifi moel θ( is arrie out by performing stohasti graient esent through bak propagation for the minimization of the reonstrution error, over all examples x (t of a training image set X, J (θ ADNT ; x (t X = x (t x (t x (t Sine the moel is being traine to reonstrut the input ata, it might en-up learning an ientity funtion an reproue the input ata. Appropriate settings in the onfigurations of the ADNT are therefore require to ensure that a lass speifi moel learns the unerlying struture of the ata an proues useful representations. For our ADNT, sine the number of noes in the first hien layer are larger than the imensions of the input ata, we first learn an overomplete representation of the ata by mapping it to a high imensional spae. This high imensional representation is then followe by a bottlenek i.e. the ata is mappe bak to a ompat, abstrat an low imensional representation in the subsequent layers of the enoer. With suh mapping, the reunant information in the ata is isare an only the require useful ontent of the ata is retaine. In orer to avoi over-fitting an improve generalization of the learnt moel to unknown test ata, we introue regularization terms into the ost funtion of ADNT. A weight eay penalty term J w an a sparsity onstraint J sp are (8

5 ae an the moifie ost funtion beomes, J reg ( θ ADNT; x (t X = x (t x (t x (t +λwj w +λ spj sp λ w an λ sp are regularization parameters. J w ensures small values of weights for all hien units. It is efine as the summation of the Frobenius norm of all weight matries, J w = 3 i W e (i + F 3 i W (i F (9 (0 J sp enfores that the mean ativation ρ (i (over all training examples of the th unit of the ith hien layer is as lose as possible to a sparsity target ρ (typially a small value, set to 0 3 in our experiments in Se. 4. It is efine in terms of the KL ivergene as, J sp = = 5 ( KL ρ ρ (i i 5 ρ log i ρ ρ (i + ( ρ log ρ ρ (i ( A lass-speifi moel θ( is ahieve by training the regularize ADNT over all images of the set X, ( θ( = min J reg θ ADNT ; x (t X ( θ ADNT A lass-speifi moel θ( is therefore mae to learn the unerlying struture of the manifol on whih fae images of that lass lie. Sine the ativation funtions use are non-linear an a number of layers are stake together, the AE struture is apable of learning very omplex non-linear manifol strutures. Classifiation: { Given a test image set X test = x (, x (, x (Ntest}, we separately reonstrut (using Eqs. & eah of its image x (t X test from all lass speifi moels θ(, = k. If x (t ( is the reonstrution of the image x (t from moel θ( (the moel finetune with images of X, then the reonstrution error is given by, r (t ( = x (t x (t ( (3 After omputing the reonstrution errors for all k moels, the eision about the lass y (t of the image x (t is mae base upon the riteria of minimum reonstrution error, y (t = arg min r (t ( (4 Here the iea is that the unseen image x (t will be reonstrute with the least error only from the moel traine from images with the same label. Following this proeure, the lass labels of all N test images of the test set are ompute. The label y test of the test image set X test is then efine as the most reurring label amongst all images of X test. This is given by, y test = arg max t δ (y (t,where { δ (y (t, y (t (5 = = 0, otherwise Algorithm Propose Image Set Classifiation Metho Input: Training ata: k Image Sets {X } k s.t. X = { x (t} N labels: y [,, k] Testing Data: Image Set X test = { x (t} N test Output: Label y test of X test Training Define ADNT struture { } Unsupervise ata: X u x (t X ; [,, k] Train GRBMs using X u to initialize θ ADNT = {θ W, θ b } for = k o ( θ( min J reg θ ADNT; x (t X θ ADNT en for Testing for eah image x (t X test o for θ( = θ( θ(k o h (t s(w e (3 s(w e ( s(w e ( x (t + b ( e + b ( e + b (3 e x (t ( s(w (3 r (t ( x (t x (t ( en for s(w( s(w( h(t + b ( + b( + b(3 Assign label to image x (t : y (t arg min en for Label of X test: y test arg max 4. Experiments r (t ( t δ(y(t. See Eq. 5 The performane of our propose metho is evaluate on four ata sets for the task of image set lassifiation for fae reognition. These atasets inlue three gray sale fae vieo atasets: Hona/UCSD ataset [8], CMU Mobo ataset [6], YouTube Celebrities ataset [6]; an an RGB-D Kinet ataset obtaine by ombining three Kinet atasets. The etaile esription of eah of these atasets an their performane evaluation using our metho an state-of-the-art methos is presente in Se 4.. Here, we first esribe the pre-proessing steps an the ommon experimental onfigurations followe for all atasets.

6 4.. Experimental Settings The fae from eah frame in the vieos of Hona/UCSD an Mobo atasets is automatially etete using Viola an Jones fae etetion algorithm [4]. It was observe that fae etetion by [4] faile in a signifiant number of frames in the ase of YouTube Celebrities ataset ue to its poor image resolution an large hea rotations. We use [] to trak the fae region aross every vieo sequene given the loation of the fae winow in the first frame (provie with the ataset. In the ase of Kinet fae atasets, the ranom regression forrest base lassifier propose in [5] is use to automatially etet faes from epth images. As epth ata is pre-aligne with RGB, the same loation of the etete fae in the epth image is use for the orresponing RGB image. After a suessful etetion, the fae region is roppe an all olore images are onverte to gray sale levels. The roppe gray sale images are then resize to 0 0, an for Hona/UCSD, Mobo an YouTube elebrities atasets respetively. The epth an the gray sale images of the Kinet atasets are resize to 0 0. Histogram equalization is applie on all images to minimize illumination variations. No other pre-proessing suh as bakgroun removal or alignment is applie. Eah roppe an histogram equalize fae image is then ivie into 4 4 (5 5 in ase of CMU Mobo ataset, as in [4, 4] istint non-overlapping uniformly spae retangular bloks an R 59 histograms of LBP8, u [0] are ompute for every blok. Histograms from all bloks are onatenate into a single vetor whih is use as a fae feature vetor in all of our experiments. In ase of the Kinet ataset, the LBP feature vetors for gray sale an epth images are onatenate an the resulting feature vetor is use. Compare Methos We ompare our propose metho with a number of reently propose state of the art image set lassifiation methos. These inlue Disriminant Canonial Correlation Analysis (DCC [7], Manifolto-Manifol Distane (MMD [7], Manifol Disriminant Analysis (MDA [5], the Linear version of the Affine Hullbase Image Set Distane (AHISD [4], the Convex Hullbase Image Set Distane (CHISD [4], Sparse Approximate Nearest Points (SANP [4], Covariane Disriminant Learning (CDL [6] an Set to Set Distane Metri Learning (SSDML [9]. The implementations provie by the respetive authors are use for all methos exept CDL whih was arefully implemente by us. The parameters for all methos are optimize for best performane. Speifially, for DCC, we set the imensions of the embeing spae to 00. The number of retaine imensions for a subspae are set to 0 (90% energy is preserve an the orresponing 0 maximum anonial orrelations are use to ompute set-set similarity. The parameters for MMD an MDA are aopte from [7] an [5] respetively. No parameter settings are require for AHISD. For CHISD, the same error penalty term (C = 00 as in [4] is aopte. For SANP, same weight parameters as in [4] are aopte for onvex optimization. No parameter settings are require for CDL an SSDML. 4.. Results an Analysis Hona/UCSD Dataset: The Hona/UCSD ataset [8] ontains 59 vieo sequenes of 0 ifferent subets. The number of frames for eah vieo sequene varies from to 645. For our experiments, we onsier eah vieo as an image set. Similar to [8, 4, 7, 5], we use 0 vieo sequenes for training an the remaining 39 for testing. In orer to ahieve onsisteny in the results, we repeat our experiments ten times with ifferent ranom seletions of the training an testing sets. The ahieve performane in terms of average ientifiation rates an stanar eviations of our metho an the ompare methos is presente in Table. The results show that the propose metho ahieves perfet lassifiation on the Hona/UCSD ata set. CMU Mobo Dataset: The Mobo (Motion of Boy ataset [6] was originally reate for human boy pose ientifiation. The ataset ontains a total of 96 sequenes of 4 subets walking on a treamill. Similar to [4, 7, 4], we ranomly selet one sequene of a subet for training an the remaining three sequenes are use for testing. We repeat our experiments 0 times for ifferent ranom seletions of the training an the testing sets. The average ientifiation rates of our propose metho along with a omparison with other methos is provie in Table. The results suggest that the propose metho ahieves a very high performane of 97.96% an outperforms the other methos. YouTube Celebrities Dataset: YouTube Celebrities [6] is the largest an the most hallenging ataset use for image set lassifiation base fae reognition. The ataset ontains 90 vieos of 47 elebrities ollete from YouTube. The fae images of the ataset exhibit a large iversity an variations in the form of pose, illumination an expressions. Moreover, the quality an resolution of the images is very low ue to the high ompression rate. Sine the fae regions in the vieos are roppe by traking [], the low image quality introues many traking errors an the region of the roppe fae is not uniform aross frames of even the same vieo. We iretly use the fae region automatially extrate from traking an o not refine its ropping by enforing onstraints as in [6]. For performane evaluation, we use five fol ross valiation experimental settings as propose in [4, 5, 5].

7 Methos Hona/UCSD CMU Mobo YouTube Kinet DCC CVPR 07 [7] 9.56 ± ± ± ±.00 MMD CVPR 08 [7] 9.05 ± ± ± ±.5 MDA CVPR 09 [5] ± ± ± ± 3.57 AHISD CVPR 0 [4] 9.8 ± ± ± ±.8 CHISD CVPR 0 [4] 93.6 ± ± ± ±.9 SANP CVPR [4] 95.3 ± ± ± ± 3. CDL CVPR [6] ± ± ± ± 0.96 SSDML ICCV 3 [9] 86.4 ± ± ± ± 3.39 Our Metho 00.0 ± ± ± ±.69 Table. Experimental Results on Hona, CMU, YouTube an Kinet atasets for ifferent methos The whole ataset is equally ivie (with minimum overlap into five fols with 9 image sets per subet in eah fol. Three of these image sets are ranomly selete for training, whereas the remaining six sets are use for testing. Table summarizes the average ientifiation rates an the stanar eviations of ifferent methos. It an be observe that the ahieve ientifiation rates for all methos are low for this ataset ompare with the Hona/UCSD an Mobo ataset. This is owing to the hallenging nature of the ataset. The vieos have been apture in real life senarios an they exhibit a wie range of appearane variations. The results suggest that our propose metho signifiantly outperforms the existing methos an ahieves a relative performane improvement of 9.0% over the seon best metho. Figure 3. Example images from gray sale atasets: Hona/UCSD (top, CMU/Mobo (enter an YouTube (bottom. Eah row orrespons to images of one ientity. Figure 4. Sample images from Kinet atasets: CurtinFaes (top, Biwi (enter an our ataset (bottom Kinet Dataset: We also evaluate the performane of our propose metho for RGB-D base fae reognition from Kinet ata. Fae reognition from Kinet ata is still in its infany an only a few work have aresse this problem [9]. The metho by Li et al. [9] first pre-proesses Kinet epth images to ahieve a anonial frontal view for faes with profile an non-frontal views. The sparse representation base lassifiation metho of [8] is use for reognition. One evaluate on CurtinFaes, the metho ahieves a lassifiation rate of 9.% for RGB, 88.7% for D an 96.7% for fusion of RGB-D ata. The propose metho is single frame base an oes not make use of the plentitue of ata whih an be instantly aquire from a Kinet sensor (30 frames per seon. Here we formulate fae reognition from Kinet ata as an RGB-D base image set lassifiation problem. Our formulation avois expensive pre-proessing steps (suh as hole filling, spike removal an anonial view estimation; otherwise require for single image base lassifiation an effetively makes use of the abunant an reaily available Kinet ata. The metho in [9] is evaluate on CurtinFaes (a Kinet RGB-D atabase of 5 subets. For our image set lassifiation experiments, we ombine three Kinet atasets: CurtinFaes [9], Biwi Kinet [5] an an in-house ataset aquire at our lab. The number of subets in eah of these atasets is 5 (5000 RGB-D images, 0 (5,000 RGB-D images an 48 (5000 RGB-D images respetively. Sample RGB images from these atasets are shown in Figure 4. Eah row orrespons to images of a person taken from CurtinFaes (top row, Biwi (mile row an our Kinet ataset (last row. These atasets are ombine into a single ataset of 0 subets. The images in the oint ataset have a large range of variations in the form of hanging illumination onitions, hea pose rotations, expression eformations, sunglass isguise, an olusions by han. For performane evaluation, RGB-D images of eah subet are ranomly ivie into five uniform fols. Consiering eah fol as an image set, we selet one set for training an the remaining sets for testing. All experiments are repeate five times for ifferent seletions of training an testing sets. The results average over five iterations are summarize in Table. The results show that the propose metho ahieves a very high performane. The results suggest that image set lassifiation proves to be a better hoie for Kinet base fae reognition. It avois omputationally expensive pre-proessing steps an the ahieve ientifiation rates with all image set lassifiation tehniques in Table are omparable or better than the single image base tehnique (96.7% of [9].

8 Aknowlegements This work is supporte by SIRF sholarship from The University of Western Australia (UWA an Australian Researh Counil (ARC grant DP0066. Thanks to Salman H. Khan for useful isussions. 5. Conlusion We propose a novel eep learning framework for image set lassifiation. An aaptive multi-layer auto-enoer struture has been introue whih is first pre-traine for appropriate parameter initialization an then use for learning lass speifi moels. A lass speifi moel automatially learns the unerlying non-linear omplex geometri surfae of the images of that lass. These learnt moels are then use for a minimum reonstrution error base lassifiation strategy uring testing. The propose framework was extensively evaluate on three benhmark gray sale atasets as well as an RGB-D Kinet ataset an state of the art performane has been ahieve. Referenes [] O. Aranelovi, G. Shakhnarovih, J. Fisher, R. Cipolla, an T. Darrell. Fae reognition with image sets using manifol ensity ivergene. In CVPR, volume, pages IEEE, 005. [] Y. Bengio, A. Courville, an P. Vinent. Representation learning: A review an new perspetives. TPAMI, 35(8:798 88, 03. [3] M. A. Carreira-Perpinan an G. E. Hinton. On ontrastive ivergene learning. In Artifiial Intelligene an Statistis, volume 005, page 7, 005. [4] H. Cevikalp an B. Triggs. Fae reognition base on image sets. In CVPR, pages IEEE, 00. [5] G. Fanelli, J. Gall, an L. Van Gool. Real time hea pose estimation with ranom regression forests. In CVPR, pages IEEE, 0. [6] R. Gross an J. Shi. The mu motion of boy (mobo atabase. Tehnial report, 00. [7] J. Hamm an D. Lee. Grassmann isriminant analysis: a unifying view on subspae-base learning. In ICML, pages ACM, 008. [8] M. Harani, C. Sanerson, S. Shirazi, an B. Lovell. Graph embeing isriminant analysis on grassmannian manifols for improve image set mathing. In CVPR, pages IEEE, 0. [9] M. T. Harani, C. Sanerson, A. Wiliem, an B. C. Lovell. Kernel analysis over riemannian manifols for visual reognition of ations, peestrians an textures. In WACV, pages IEEE, 0. [0] M. Hayat, M. Bennamoun, an A. A. El-Sallam. Clustering of vieo-pathes on grassmannian manifol for faial expression reognition from 3 vieos. In WACV, 03. [] G. Hinton, S. Osinero, M. Welling, an Y.-W. Teh. Unsupervise isovery of nonlinear struture using ontrastive bakpropagation. Cognitive siene, 30(4:75 73, 006. [] G. E. Hinton, S. Osinero, an Y.-W. Teh. A fast learning algorithm for eep belief nets. Neural omputation, 8(7:57 554, 006. [3] G. E. Hinton an R. R. Salakhutinov. Reuing the imensionality of ata with neural networks. Siene, 33(5786: , 006. [4] Y. Hu, A. S. Mian, an R. Owens. Sparse approximate nearest points for image set lassifiation. In CVPR, pages 8. IEEE, 0. [5] S. H. Khan, M. Bennamoun, F. Sohel, an R. Togneri. Automati feature learning for robust shaow etetion. In CVPR, 04. [6] M. Kim, S. Kumar, V. Pavlovi, an H. Rowley. Fae traking an reognition with visual onstraints in real-worl vieos. In CVPR, pages 8. IEEE, 008. [7] T. Kim, J. Kittler, an R. Cipolla. Disriminative learning an reognition of image set lasses using anonial orrelations. IEEE TPAMI, 9(6:005 08, 007. [8] K.-C. Lee, J. Ho, M.-H. Yang, an D. Kriegman. Vieobase fae reognition using probabilisti appearane manifols. In CVPR, volume, pages I 33. IEEE, 003. [9] B. Y. Li, A. S. Mian, W. Liu, an A. Krishna. Using kinet for fae reognition uner varying poses, expressions, illumination an isguise. In WACV, pages IEEE, 03. [0] T. Oala, M. Pietikainen, an T. Maenpaa. Multiresolution gray-sale an rotation invariant texture lassifiation with loal binary patterns. TPAMI, 4(7:97 987, 00. [] D. A. Ross, J. Lim, R.-S. Lin, an M.-H. Yang. Inremental learning for robust visual traking. IJCV, 77(-3:5 4, 008. [] P. Turaga, A. Veeraraghavan, A. Srivastava, an R. Chellappa. Statistial omputations on grassmann an stiefel manifols for image an vieo-base reognition. IEEE TPAMI, 33(:73 86, nov. 0. [3] M. Uzair, A. Mahmoo, A. Mian, an C. MDonal. A ompat isriminative representation for effiient image-set lassifiation with appliation to biometri reognition. In ICB, pages 8, 03. [4] P. Viola an M. J. Jones. Robust real-time fae etetion. IJCV, 57:37 54, 004. [5] R. Wang an X. Chen. Manifol isriminant analysis. In CVPR, pages IEEE, 009. [6] R. Wang, H. Guo, L. S. Davis, an Q. Dai. Covariane isriminative learning: A natural an effiient approah to image set lassifiation. In CVPR, pages IEEE, 0. [7] R. Wang, S. Shan, X. Chen, an W. Gao. Manifol-manifol istane with appliation to fae reognition base on image set. In CVPR, pages 8. IEEE, 008. [8] J. Wright, A. Y. Yang, A. Ganesh, S. S. Sastry, an Y. Ma. Robust fae reognition via sparse representation. TPAMI, 3:0 7, 009. [9] P. Zhu, L. Zhang, W. Zuo, an D. Zhang. From point to set: Exten the learning of istane metris. In ICCV, 03.

Image Set Classification Based on Synthetic Examples and Reverse Training

Image Set Classification Based on Synthetic Examples and Reverse Training Image Set Classification Based on Synthetic Examples and Reverse Training Qingjun Liang 1, Lin Zhang 1(&), Hongyu Li 1, and Jianwei Lu 1,2 1 School of Software Engineering, Tongji University, Shanghai,

More information

LAB 4: Operations on binary images Histograms and color tables

LAB 4: Operations on binary images Histograms and color tables LAB 4: Operations on binary images Histograms an olor tables Computer Vision Laboratory Linköping University, Sween Preparations an start of the lab system You will fin a ouple of home exerises (marke

More information

KERNEL SPARSE REPRESENTATION WITH LOCAL PATTERNS FOR FACE RECOGNITION

KERNEL SPARSE REPRESENTATION WITH LOCAL PATTERNS FOR FACE RECOGNITION KERNEL SPARSE REPRESENTATION WITH LOCAL PATTERNS FOR FACE RECOGNITION Cuiui Kang 1, Shengai Liao, Shiming Xiang 1, Chunhong Pan 1 1 National Laboratory of Pattern Reognition, Institute of Automation, Chinese

More information

Learning Spatiotemporal Features for Infrared Action Recognition with 3D Convolutional Neural Networks

Learning Spatiotemporal Features for Infrared Action Recognition with 3D Convolutional Neural Networks Learning Spatiotemporal Features for Infrare Ation Reognition with 3D Convolutional Neural Networks Zhuolin Jiang, Viktor Rozgi, Sanar Aali Raytheon BBN Tehnologies, Cambrige, MA, 02138 {zjiang, vrozgi,

More information

Unsupervised Segmentation of Stereoscopic Video Objects: Proposal. and Comparison of Two Depth-Based Approaches

Unsupervised Segmentation of Stereoscopic Video Objects: Proposal. and Comparison of Two Depth-Based Approaches Unsupervise Segmentation of Stereosopi Vieo Objets: Proposal an Comparison of Two Depth-Base Approahes Klimis S. Ntalianis an Athanasios S.Drigas Net Meia Lab, NCSR Demokritos, Athens, Greee E-mail: kntal@image.ntua.gr

More information

Learning Convention Propagation in BeerAdvocate Reviews from a etwork Perspective. Abstract

Learning Convention Propagation in BeerAdvocate Reviews from a etwork Perspective. Abstract CS 9 Projet Final Report: Learning Convention Propagation in BeerAdvoate Reviews from a etwork Perspetive Abstrat We look at the way onventions propagate between reviews on the BeerAdvoate dataset, and

More information

A Coarse-to-Fine Classification Scheme for Facial Expression Recognition

A Coarse-to-Fine Classification Scheme for Facial Expression Recognition A Coarse-to-Fine Classifiation Sheme for Faial Expression Reognition Xiaoyi Feng 1,, Abdenour Hadid 1 and Matti Pietikäinen 1 1 Mahine Vision Group Infoteh Oulu and Dept. of Eletrial and Information Engineering

More information

Multitarget Data Association with Higher-Order Motion Models

Multitarget Data Association with Higher-Order Motion Models Multitarget Data Assoiation with Higher-Orer Motion Moels Robert T. Collins The Pennsylvania State University University Park, PA 16802, USA Abstrat We present an iterative approximate solution to the

More information

the data. Structured Principal Component Analysis (SPCA)

the data. Structured Principal Component Analysis (SPCA) Strutured Prinipal Component Analysis Kristin M. Branson and Sameer Agarwal Department of Computer Siene and Engineering University of California, San Diego La Jolla, CA 9193-114 Abstrat Many tasks involving

More information

NONLINEAR BACK PROJECTION FOR TOMOGRAPHIC IMAGE RECONSTRUCTION. Ken Sauer and Charles A. Bouman

NONLINEAR BACK PROJECTION FOR TOMOGRAPHIC IMAGE RECONSTRUCTION. Ken Sauer and Charles A. Bouman NONLINEAR BACK PROJECTION FOR TOMOGRAPHIC IMAGE RECONSTRUCTION Ken Sauer and Charles A. Bouman Department of Eletrial Engineering, University of Notre Dame Notre Dame, IN 46556, (219) 631-6999 Shool of

More information

The Minimum Redundancy Maximum Relevance Approach to Building Sparse Support Vector Machines

The Minimum Redundancy Maximum Relevance Approach to Building Sparse Support Vector Machines The Minimum Redundany Maximum Relevane Approah to Building Sparse Support Vetor Mahines Xiaoxing Yang, Ke Tang, and Xin Yao, Nature Inspired Computation and Appliations Laboratory (NICAL), Shool of Computer

More information

Solutions to Tutorial 2 (Week 9)

Solutions to Tutorial 2 (Week 9) The University of Syney Shool of Mathematis an Statistis Solutions to Tutorial (Week 9) MATH09/99: Disrete Mathematis an Graph Theory Semester, 0. Determine whether eah of the following sequenes is the

More information

arxiv: v1 [cs.cv] 10 Jan 2017

arxiv: v1 [cs.cv] 10 Jan 2017 Efficient Image Set Classification using Linear Regression based Image Reconstruction S.A.A. Shah, U. Nadeem, M. Bennamoun, F. Sohel, and R. Togneri arxiv:1701.02485v1 [cs.cv] 10 Jan 2017 The University

More information

An Efficient Image Distortion Correction Method for an X-ray Digital Tomosynthesis System

An Efficient Image Distortion Correction Method for an X-ray Digital Tomosynthesis System An Effiient Image Distortion Corretion Metho for an X-ray Digital Tomosynthesis System J.Y. Kim Dept. of Mehatronis Engineering, Tongmyong University of Information Tehnology, 55 Yongang-ong, Nam-gu, Busan

More information

Boosted Random Forest

Boosted Random Forest Boosted Random Forest Yohei Mishina, Masamitsu suhiya and Hironobu Fujiyoshi Department of Computer Siene, Chubu University, 1200 Matsumoto-ho, Kasugai, Aihi, Japan {mishi, mtdoll}@vision.s.hubu.a.jp,

More information

Evolutionary Feature Synthesis for Image Databases

Evolutionary Feature Synthesis for Image Databases Evolutionary Feature Synthesis for Image Databases Anlei Dong, Bir Bhanu, Yingqiang Lin Center for Researh in Intelligent Systems University of California, Riverside, California 92521, USA {adong, bhanu,

More information

A Novel Validity Index for Determination of the Optimal Number of Clusters

A Novel Validity Index for Determination of the Optimal Number of Clusters IEICE TRANS. INF. & SYST., VOL.E84 D, NO.2 FEBRUARY 2001 281 LETTER A Novel Validity Index for Determination of the Optimal Number of Clusters Do-Jong KIM, Yong-Woon PARK, and Dong-Jo PARK, Nonmembers

More information

Predicting Project Outcome Leveraging Socio-Technical Network Patterns

Predicting Project Outcome Leveraging Socio-Technical Network Patterns 203 7th European Conferene on Software Maintenane an Reengineering Preiting Projet Outome Leveraging Soio-Tehnial Network Patterns Dii Surian, Yuan Tian, Davi Lo, Hong Cheng an Ee-Peng Lim Shool of Information

More information

Empowering Simple Binary Classifiers for Image Set based Face Recognition

Empowering Simple Binary Classifiers for Image Set based Face Recognition Noname manuscript No. (will be inserted by the editor) Empowering Simple Binary Classifiers for Image Set based Face Recognition Munawar Hayat Salman H. Khan Mohammed Bennamoun Received: date / Accepted:

More information

Patch Volumes: Segmentation-based Consistent Mapping with RGB-D Cameras

Patch Volumes: Segmentation-based Consistent Mapping with RGB-D Cameras Path Volumes: Segmentation-base Consistent Mapping with RGB-D Cameras Peter Henry an Dieter Fox University of Washington Computer Siene & Engineering Seattle, Washington peter@s.washington.eu fox@s.washington.eu

More information

arxiv: v2 [cs.cv] 3 Mar 2019

arxiv: v2 [cs.cv] 3 Mar 2019 Real Time Surveillance for Low Resolution and Limited-Data Scenarios: An Image Set Classification Approach arxiv:1803.09470v2 [cs.cv] 3 Mar 2019 Uzair Nadeem Syed Afaq Ali Shah Mohammed Bennamoun Roberto

More information

Anchoring quartet-based phylogenetic distances and applications to species tree reconstruction

Anchoring quartet-based phylogenetic distances and applications to species tree reconstruction Anhoring quartet-base phylogeneti istanes an appliations to speies tree reonstrution Erfan Sayyari an Siavash Mirarab Department of Eletrial an Computer Engineering University of California at San Diego

More information

Capturing Large Intra-class Variations of Biometric Data by Template Co-updating

Capturing Large Intra-class Variations of Biometric Data by Template Co-updating Capturing Large Intra-lass Variations of Biometri Data by Template Co-updating Ajita Rattani University of Cagliari Piazza d'armi, Cagliari, Italy ajita.rattani@diee.unia.it Gian Lua Marialis University

More information

Performance of Histogram-Based Skin Colour Segmentation for Arms Detection in Human Motion Analysis Application

Performance of Histogram-Based Skin Colour Segmentation for Arms Detection in Human Motion Analysis Application World Aademy of Siene, Engineering and Tehnology 8 009 Performane of Histogram-Based Skin Colour Segmentation for Arms Detetion in Human Motion Analysis Appliation Rosalyn R. Porle, Ali Chekima, Farrah

More information

Learning Discriminative and Shareable Features. Scene Classificsion

Learning Discriminative and Shareable Features. Scene Classificsion Learning Disriminative and Shareable Features for Sene Classifiation Zhen Zuo, Gang Wang, Bing Shuai, Lifan Zhao, Qingxiong Yang, and Xudong Jiang Nanyang Tehnologial University, Singapore, Advaned Digital

More information

Optimization of Image Processing in Video-based Traffic Monitoring

Optimization of Image Processing in Video-based Traffic Monitoring http://x.oi.org/0.5755/j0.eee.8.8.634 Optimization of Image Proessing in Vieo-base Traffi Monitoring Fei Zhu, Jiamin Ning, Yong Ren, Jingyu Peng Shool of Computer Siene an Tehnology, Soohow University,

More information

Abstract. Key Words: Image Filters, Fuzzy Filters, Order Statistics Filters, Rank Ordered Mean Filters, Channel Noise. 1.

Abstract. Key Words: Image Filters, Fuzzy Filters, Order Statistics Filters, Rank Ordered Mean Filters, Channel Noise. 1. Fuzzy Weighted Rank Ordered Mean (FWROM) Filters for Mixed Noise Suppression from Images S. Meher, G. Panda, B. Majhi 3, M.R. Meher 4,,4 Department of Eletronis and I.E., National Institute of Tehnology,

More information

Pipelined Multipliers for Reconfigurable Hardware

Pipelined Multipliers for Reconfigurable Hardware Pipelined Multipliers for Reonfigurable Hardware Mithell J. Myjak and José G. Delgado-Frias Shool of Eletrial Engineering and Computer Siene, Washington State University Pullman, WA 99164-2752 USA {mmyjak,

More information

Detection and Recognition of Non-Occluded Objects using Signature Map

Detection and Recognition of Non-Occluded Objects using Signature Map 6th WSEAS International Conferene on CIRCUITS, SYSTEMS, ELECTRONICS,CONTROL & SIGNAL PROCESSING, Cairo, Egypt, De 9-31, 007 65 Detetion and Reognition of Non-Oluded Objets using Signature Map Sangbum Park,

More information

arxiv: v3 [cs.cv] 1 Apr 2015

arxiv: v3 [cs.cv] 1 Apr 2015 Representation Learning with Deep Extreme Learning Machines for Efficient Image Set Classification Muhammad Uzair 1, Faisal Shafait 1, Bernard Ghanem 2 and Ajmal Mian 1 1 Computer Science & Software Engineering,

More information

Improved Vehicle Classification in Long Traffic Video by Cooperating Tracker and Classifier Modules

Improved Vehicle Classification in Long Traffic Video by Cooperating Tracker and Classifier Modules Improved Vehile Classifiation in Long Traffi Video by Cooperating Traker and Classifier Modules Brendan Morris and Mohan Trivedi University of California, San Diego San Diego, CA 92093 {b1morris, trivedi}@usd.edu

More information

Figure 1. LBP in the field of texture analysis operators.

Figure 1. LBP in the field of texture analysis operators. L MEHODOLOGY he loal inary pattern (L) texture analysis operator is defined as a gray-sale invariant texture measure, derived from a general definition of texture in a loal neighorhood. he urrent form

More information

Unsupervised Stereoscopic Video Object Segmentation Based on Active Contours and Retrainable Neural Networks

Unsupervised Stereoscopic Video Object Segmentation Based on Active Contours and Retrainable Neural Networks Unsupervised Stereosopi Video Objet Segmentation Based on Ative Contours and Retrainable Neural Networks KLIMIS NTALIANIS, ANASTASIOS DOULAMIS, and NIKOLAOS DOULAMIS National Tehnial University of Athens

More information

Conflicts Analysis for Inter-Enterprise Business Process Model

Conflicts Analysis for Inter-Enterprise Business Process Model Conflits Analysis for nter-enterprise Business Proess Moel Wei DNG, Zhong TAN, Jian WANG, Jun ZHU, Haiqi LANG,Lei ZHANG {ingw, tianz, wangwj, zhujun, lianghq, lzhang}@n.ibm.om BM China Researh Lab, BM

More information

The Thresholding MLEM Algorithm

The Thresholding MLEM Algorithm Journal of Meial an Biologial Engineering, 24(2: 85-9 85 The Thresholing MLEM Algorithm Keh-Shih Chuang, Meei-Ling Jan,2 Jay Wu Sharon Chen Yu-Ching Ni Ying-Kai Fu 2 epartment of Nulear Siene, National

More information

Generalized Buffering of PTL Logic Stages using Boolean Division and Don t Cares

Generalized Buffering of PTL Logic Stages using Boolean Division and Don t Cares Generalize Buffering of PTL Logi Stages using Boolean Division an Don t Cares Rajesh Garg Sunil P Khatri Department of ECE, Texas A&M University, College Station, TX 77843 Abstrat Pass Transistor Logi

More information

DOMAIN ADAPTATION BY ITERATIVE IMPROVEMENT OF SOFT-LABELING AND MAXIMIZATION OF NON-PARAMETRIC MUTUAL INFORMATION. M.N.A. Khan, Douglas R.

DOMAIN ADAPTATION BY ITERATIVE IMPROVEMENT OF SOFT-LABELING AND MAXIMIZATION OF NON-PARAMETRIC MUTUAL INFORMATION. M.N.A. Khan, Douglas R. DOMAIN ADAPTATION BY ITERATIVE IMPROVEMENT OF SOFT-LABELING AND MAXIMIZATION OF NON-PARAMETRIC MUTUAL INFORMATION M.N.A. Khan, Douglas R. Heisterkamp Department of Computer Siene Oklahoma State University,

More information

IN the recent years, due to the wide availability

IN the recent years, due to the wide availability 1 Image Set Classification for Low Resolution Surveillance Uzair Nadeem, Syed Afaq Ali Shah, Mohammed Bennamoun, Roberto Togneri and Ferdous Sohel arxiv:1803.09470v1 [cs.cv] 26 Mar 2018 Abstract This paper

More information

New Fuzzy Object Segmentation Algorithm for Video Sequences *

New Fuzzy Object Segmentation Algorithm for Video Sequences * JOURNAL OF INFORMATION SCIENCE AND ENGINEERING 24, 521-537 (2008) New Fuzzy Obet Segmentation Algorithm for Video Sequenes * KUO-LIANG CHUNG, SHIH-WEI YU, HSUEH-JU YEH, YONG-HUAI HUANG AND TA-JEN YAO Department

More information

Multi-Piece Mold Design Based on Linear Mixed-Integer Program Toward Guaranteed Optimality

Multi-Piece Mold Design Based on Linear Mixed-Integer Program Toward Guaranteed Optimality INTERNATIONAL CONFERENCE ON MANUFACTURING AUTOMATION (ICMA200) Multi-Piee Mold Design Based on Linear Mixed-Integer Program Toward Guaranteed Optimality Stephen Stoyan, Yong Chen* Epstein Department of

More information

Efficient Image Set Classification using Linear Regression based Image Reconstruction

Efficient Image Set Classification using Linear Regression based Image Reconstruction Efficient Image Set Classification using Linear Regression based Image Reconstruction S.A.A. Shah, U. Nadeem, M. Bennamoun, F. Sohel, and R. Togneri The University of Western Australia Murdoch University

More information

Cluster-Based Cumulative Ensembles

Cluster-Based Cumulative Ensembles Cluster-Based Cumulative Ensembles Hanan G. Ayad and Mohamed S. Kamel Pattern Analysis and Mahine Intelligene Lab, Eletrial and Computer Engineering, University of Waterloo, Waterloo, Ontario N2L 3G1,

More information

Discrete sequential models and CRFs. 1 Case Study: Supervised Part-of-Speech Tagging

Discrete sequential models and CRFs. 1 Case Study: Supervised Part-of-Speech Tagging 0-708: Probabilisti Graphial Models 0-708, Spring 204 Disrete sequential models and CRFs Leturer: Eri P. Xing Sribes: Pankesh Bamotra, Xuanhong Li Case Study: Supervised Part-of-Speeh Tagging The supervised

More information

Exploiting Enriched Contextual Information for Mobile App Classification

Exploiting Enriched Contextual Information for Mobile App Classification Exploiting Enrihed Contextual Information for Mobile App Classifiation Hengshu Zhu 1 Huanhuan Cao 2 Enhong Chen 1 Hui Xiong 3 Jilei Tian 2 1 University of Siene and Tehnology of China 2 Nokia Researh Center

More information

Shape Outlier Detection Using Pose Preserving Dynamic Shape Models

Shape Outlier Detection Using Pose Preserving Dynamic Shape Models Shape Outlier Detetion Using Pose Preserving Dynami Shape Models Chan-Su Lee Ahmed Elgammal Department of Computer Siene, Rutgers University, Pisataway, NJ 8854 USA CHANSU@CS.RUTGERS.EDU ELGAMMAL@CS.RUTGERS.EDU

More information

Dynamic Restoration in Multi-layer IP/MPLS-over- Flexgrid Networks

Dynamic Restoration in Multi-layer IP/MPLS-over- Flexgrid Networks Dynami Restoration in Multi-layer IP/MPLS-over- Flexgri Networks Alberto Castro, Luis Velaso, Jaume Comellas, an Gabriel Junyent Universitat Politènia e Catalunya (UPC), Barelona, Spain E-mail: aastro@a.up.eu

More information

Inverse Design of Urban Procedural Models

Inverse Design of Urban Procedural Models Inverse Design of Uran Proeural Moels Carlos A. Vanegas Purue University U.C. Berkeley Ignaio Garia-Dorao Purue University Daniel G. Aliaga Purue University Development site Paul Waell U.C. Berkeley Non-optimize

More information

Polyhedron Volume-Ratio-based Classification for Image Recognition

Polyhedron Volume-Ratio-based Classification for Image Recognition Polyhedron Volume-Ratio-based Classifiation for Image Reognition Qingxiang Feng, Jeng-Shyang Pan, Senior Member, IEEE, Jar-Ferr Yang, Fellow, IEEE and Yang-ing Chou Abstrat In this paper, a novel method,

More information

Particle Swarm Optimization for the Design of High Diffraction Efficient Holographic Grating

Particle Swarm Optimization for the Design of High Diffraction Efficient Holographic Grating Original Artile Partile Swarm Optimization for the Design of High Diffration Effiient Holographi Grating A.K. Tripathy 1, S.K. Das, M. Sundaray 3 and S.K. Tripathy* 4 1, Department of Computer Siene, Berhampur

More information

LARGE-SCALE INVERSE MICROWAVE BACKSCATTER MODELING OF SEA ICE

LARGE-SCALE INVERSE MICROWAVE BACKSCATTER MODELING OF SEA ICE LARGE-SCALE INVERSE MICROWAVE BACKSCATTER MODELING OF SEA ICE Quinn P Remun Mirowave Earth Remote Sensing Laboratory Brigham Young University Provo Utah Abstrat Polar sea ie harateristi provie important

More information

Algorithms to Accelerate Multiple Regular Expressions Matching for Deep Packet Inspection

Algorithms to Accelerate Multiple Regular Expressions Matching for Deep Packet Inspection Algorithms to Aelerate Multiple Regular Expressions Mathing for Deep Paket Inspetion Sailesh Kumar Washington University Computer Siene an Engineering St. Louis, MO 60-899 +--9-06 sailesh@arl.wustl.eu

More information

A Full-Featured, Error Resilient, Scalable Wavelet Video Codec Based on the Set Partitioning in Hierarchical Trees (SPIHT) Algorithm

A Full-Featured, Error Resilient, Scalable Wavelet Video Codec Based on the Set Partitioning in Hierarchical Trees (SPIHT) Algorithm A Full-Feature, Error Resilient, Salable Wavelet Vieo Coe Base on the Set Partitioning in Hierarhial Trees (SPIHT) Algorithm Sungae Cho an William A. Pearlman Center for Next Generation Vieo Researh Rensselaer

More information

Video Data and Sonar Data: Real World Data Fusion Example

Video Data and Sonar Data: Real World Data Fusion Example 14th International Conferene on Information Fusion Chiago, Illinois, USA, July 5-8, 2011 Video Data and Sonar Data: Real World Data Fusion Example David W. Krout Applied Physis Lab dkrout@apl.washington.edu

More information

Transition Detection Using Hilbert Transform and Texture Features

Transition Detection Using Hilbert Transform and Texture Features Amerian Journal of Signal Proessing 1, (): 35-4 DOI: 1.593/.asp.1.6 Transition Detetion Using Hilbert Transform and Texture Features G. G. Lashmi Priya *, S. Domni Department of Computer Appliations, National

More information

EASY TRANSFER LEARNING BY EXPLOITING INTRA-DOMAIN STRUCTURES

EASY TRANSFER LEARNING BY EXPLOITING INTRA-DOMAIN STRUCTURES EASY TRANSFER LEARNING BY EXPLOITING INTRA-DOMAIN STRUCTURES Jindong Wang 1, Yiqiang Chen 1,, Han Yu 2, Meiyu Huang 3, Qiang Yang 4 1 Beiing Key Lab. of Mobile Computing and Pervasive Devie, Inst. of Computing

More information

Image Set-based Face Recognition: A Local Multi-Keypoint Descriptor-based Approach

Image Set-based Face Recognition: A Local Multi-Keypoint Descriptor-based Approach 2013 IEEE Conference on Computer Vision and Pattern Recognition Workshops Image Set-based Face Recognition: A Local Multi-Keypoint Descriptor-based Approach Na Liu 1, Meng-Hui Lim 2, Pong C. Yuen 2, and

More information

Advanced Algorithms for Fast and Scalable Deep Packet Inspection

Advanced Algorithms for Fast and Scalable Deep Packet Inspection Avane Algorithms for Fast an Salale Deep Paket Inspetion Sailesh Kumar Washington University sailesh@arl.wustl.eu Jonathan Turner Washington University jon.turner@wustl.eu John Williams Ciso Systems jwill@iso.om

More information

Automatic Physical Design Tuning: Workload as a Sequence Sanjay Agrawal Microsoft Research One Microsoft Way Redmond, WA, USA +1-(425)

Automatic Physical Design Tuning: Workload as a Sequence Sanjay Agrawal Microsoft Research One Microsoft Way Redmond, WA, USA +1-(425) Automati Physial Design Tuning: Workload as a Sequene Sanjay Agrawal Mirosoft Researh One Mirosoft Way Redmond, WA, USA +1-(425) 75-357 sagrawal@mirosoft.om Eri Chu * Computer Sienes Department University

More information

arxiv: v2 [cs.hc] 28 Apr 2018

arxiv: v2 [cs.hc] 28 Apr 2018 Clustrophile 2: Guie Visual Clustering Analysis Maro Cavallo an Çağatay Demiralp a arxiv:1804.03048v2 [s.hc] 28 Apr 2018 Fig. 1: Clustrophile 2 is an interative tool for guie exploratory lustering analysis.

More information

Rotation Invariant Spherical Harmonic Representation of 3D Shape Descriptors

Rotation Invariant Spherical Harmonic Representation of 3D Shape Descriptors Eurographis Symposium on Geometry Proessing (003) L. Kobbelt, P. Shröder, H. Hoppe (Editors) Rotation Invariant Spherial Harmoni Representation of 3D Shape Desriptors Mihael Kazhdan, Thomas Funkhouser,

More information

arxiv: v1 [cs.db] 13 Sep 2017

arxiv: v1 [cs.db] 13 Sep 2017 An effiient lustering algorithm from the measure of loal Gaussian distribution Yuan-Yen Tai (Dated: May 27, 2018) In this paper, I will introdue a fast and novel lustering algorithm based on Gaussian distribution

More information

HEXA: Compact Data Structures for Faster Packet Processing

HEXA: Compact Data Structures for Faster Packet Processing Washington University in St. Louis Washington University Open Sholarship All Computer Siene and Engineering Researh Computer Siene and Engineering Report Number: 27-26 27 HEXA: Compat Data Strutures for

More information

Extracting Partition Statistics from Semistructured Data

Extracting Partition Statistics from Semistructured Data Extrating Partition Statistis from Semistrutured Data John N. Wilson Rihard Gourlay Robert Japp Mathias Neumüller Department of Computer and Information Sienes University of Strathlyde, Glasgow, UK {jnw,rsg,rpj,mathias}@is.strath.a.uk

More information

Fast Rigid Motion Segmentation via Incrementally-Complex Local Models

Fast Rigid Motion Segmentation via Incrementally-Complex Local Models Fast Rigid Motion Segmentation via Inrementally-Complex Loal Models Fernando Flores-Mangas Allan D. Jepson Department of Computer Siene, University of Toronto {mangas,jepson}@s.toronto.edu Abstrat The

More information

Semi-Supervised Affinity Propagation with Instance-Level Constraints

Semi-Supervised Affinity Propagation with Instance-Level Constraints Semi-Supervised Affinity Propagation with Instane-Level Constraints Inmar E. Givoni, Brendan J. Frey Probabilisti and Statistial Inferene Group University of Toronto 10 King s College Road, Toronto, Ontario,

More information

FOREGROUND OBJECT EXTRACTION USING FUZZY C MEANS WITH BIT-PLANE SLICING AND OPTICAL FLOW

FOREGROUND OBJECT EXTRACTION USING FUZZY C MEANS WITH BIT-PLANE SLICING AND OPTICAL FLOW FOREGROUND OBJECT EXTRACTION USING FUZZY C EANS WITH BIT-PLANE SLICING AND OPTICAL FLOW SIVAGAI., REVATHI.T, JEGANATHAN.L 3 APSG, SCSE, VIT University, Chennai, India JRF, DST, Dehi, India. 3 Professor,

More information

Weak Dependence on Initialization in Mixture of Linear Regressions

Weak Dependence on Initialization in Mixture of Linear Regressions Proeedings of the International MultiConferene of Engineers and Computer Sientists 8 Vol I IMECS 8, Marh -6, 8, Hong Kong Weak Dependene on Initialization in Mixture of Linear Regressions Ryohei Nakano

More information

A {k, n}-secret Sharing Scheme for Color Images

A {k, n}-secret Sharing Scheme for Color Images A {k, n}-seret Sharing Sheme for Color Images Rastislav Luka, Konstantinos N. Plataniotis, and Anastasios N. Venetsanopoulos The Edward S. Rogers Sr. Dept. of Eletrial and Computer Engineering, University

More information

Face and Facial Feature Tracking for Natural Human-Computer Interface

Face and Facial Feature Tracking for Natural Human-Computer Interface Fae and Faial Feature Traking for Natural Human-Computer Interfae Vladimir Vezhnevets Graphis & Media Laboratory, Dept. of Applied Mathematis and Computer Siene of Mosow State University Mosow, Russia

More information

Facility Location: Distributed Approximation

Facility Location: Distributed Approximation Faility Loation: Distributed Approximation Thomas Mosibroda Roger Wattenhofer Distributed Computing Group PODC 2005 Where to plae ahes in the Internet? A distributed appliation that has to dynamially plae

More information

Outline: Software Design

Outline: Software Design Outline: Software Design. Goals History of software design ideas Design priniples Design methods Life belt or leg iron? (Budgen) Copyright Nany Leveson, Sept. 1999 A Little History... At first, struggling

More information

INTRODUCTION. USB Connection Cable included

INTRODUCTION. USB Connection Cable included INTRODUCTION USB Connetion Cable inlue PRG007 is the new programming interfae for alarms, l, parking now available an future proution. It s Winows sw base on. After owloaing an installing the software

More information

Graph-Based vs Depth-Based Data Representation for Multiview Images

Graph-Based vs Depth-Based Data Representation for Multiview Images Graph-Based vs Depth-Based Data Representation for Multiview Images Thomas Maugey, Antonio Ortega, Pasal Frossard Signal Proessing Laboratory (LTS), Eole Polytehnique Fédérale de Lausanne (EPFL) Email:

More information

Relevance for Computer Vision

Relevance for Computer Vision The Geometry of ROC Spae: Understanding Mahine Learning Metris through ROC Isometris, by Peter A. Flah International Conferene on Mahine Learning (ICML-23) http://www.s.bris.a.uk/publiations/papers/74.pdf

More information

Contour Box: Rejecting Object Proposals Without Explicit Closed Contours

Contour Box: Rejecting Object Proposals Without Explicit Closed Contours Contour Box: Rejeting Objet Proposals Without Expliit Closed Contours Cewu Lu, Shu Liu Jiaya Jia Chi-Keung Tang The Hong Kong University of Siene and Tehnology Stanford University The Chinese University

More information

Detecting Outliers in High-Dimensional Datasets with Mixed Attributes

Detecting Outliers in High-Dimensional Datasets with Mixed Attributes Deteting Outliers in High-Dimensional Datasets with Mixed Attributes A. Koufakou, M. Georgiopoulos, and G.C. Anagnostopoulos 2 Shool of EECS, University of Central Florida, Orlando, FL, USA 2 Dept. of

More information

One Against One or One Against All : Which One is Better for Handwriting Recognition with SVMs?

One Against One or One Against All : Which One is Better for Handwriting Recognition with SVMs? One Against One or One Against All : Whih One is Better for Handwriting Reognition with SVMs? Jonathan Milgram, Mohamed Cheriet, Robert Sabourin To ite this version: Jonathan Milgram, Mohamed Cheriet,

More information

Spatial-Aware Collaborative Representation for Hyperspectral Remote Sensing Image Classification

Spatial-Aware Collaborative Representation for Hyperspectral Remote Sensing Image Classification Spatial-Aware Collaborative Representation for Hyperspetral Remote Sensing Image ifiation Junjun Jiang, Member, IEEE, Chen Chen, Member, IEEE, Yi Yu, Xinwei Jiang, and Jiayi Ma Member, IEEE Representation-residual

More information

TUMOR DETECTION IN MRI BRAIN IMAGE SEGMENTATION USING PHASE CONGRUENCY MODIFIED FUZZY C MEAN ALGORITHM

TUMOR DETECTION IN MRI BRAIN IMAGE SEGMENTATION USING PHASE CONGRUENCY MODIFIED FUZZY C MEAN ALGORITHM TUMOR DETECTION IN MRI BRAIN IMAGE SEGMENTATION USING PHASE CONGRUENCY MODIFIED FUZZY C MEAN ALGORITHM M. Murugeswari 1, M.Gayathri 2 1 Assoiate Professor, 2 PG Sholar 1,2 K.L.N College of Information

More information

What are Cycle-Stealing Systems Good For? A Detailed Performance Model Case Study

What are Cycle-Stealing Systems Good For? A Detailed Performance Model Case Study What are Cyle-Stealing Systems Good For? A Detailed Performane Model Case Study Wayne Kelly and Jiro Sumitomo Queensland University of Tehnology, Australia {w.kelly, j.sumitomo}@qut.edu.au Abstrat The

More information

Gradient based progressive probabilistic Hough transform

Gradient based progressive probabilistic Hough transform Gradient based progressive probabilisti Hough transform C.Galambos, J.Kittler and J.Matas Abstrat: The authors look at the benefits of exploiting gradient information to enhane the progressive probabilisti

More information

Volume 3, Issue 9, September 2013 International Journal of Advanced Research in Computer Science and Software Engineering

Volume 3, Issue 9, September 2013 International Journal of Advanced Research in Computer Science and Software Engineering Volume 3, Issue 9, September 2013 ISSN: 2277 128X International Journal of Advaned Researh in Computer Siene and Software Engineering Researh Paper Available online at: www.ijarsse.om A New-Fangled Algorithm

More information

Self-Adaptive Parent to Mean-Centric Recombination for Real-Parameter Optimization

Self-Adaptive Parent to Mean-Centric Recombination for Real-Parameter Optimization Self-Adaptive Parent to Mean-Centri Reombination for Real-Parameter Optimization Kalyanmoy Deb and Himanshu Jain Department of Mehanial Engineering Indian Institute of Tehnology Kanpur Kanpur, PIN 86 {deb,hjain}@iitk.a.in

More information

Superpixel Tracking. School of Information and Communication Engineering, Dalian University of Technology, China 2

Superpixel Tracking. School of Information and Communication Engineering, Dalian University of Technology, China 2 Superpixel Traking Shu Wang1, Huhuan Lu1, Fan Yang1, and Ming-Hsuan Yang2 1 Shool of Information and Communiation Engineering, Dalian University of Tehnology, China 2 Eletrial Engineering and Computer

More information

Contents Contents...I List of Tables...VIII List of Figures...IX 1. Introduction Information Retrieval... 8

Contents Contents...I List of Tables...VIII List of Figures...IX 1. Introduction Information Retrieval... 8 Contents Contents...I List of Tables...VIII List of Figures...IX 1. Introdution... 1 1.1. Internet Information...2 1.2. Internet Information Retrieval...3 1.2.1. Doument Indexing...4 1.2.2. Doument Retrieval...4

More information

System-Level Parallelism and Throughput Optimization in Designing Reconfigurable Computing Applications

System-Level Parallelism and Throughput Optimization in Designing Reconfigurable Computing Applications System-Level Parallelism and hroughput Optimization in Designing Reonfigurable Computing Appliations Esam El-Araby 1, Mohamed aher 1, Kris Gaj 2, arek El-Ghazawi 1, David Caliga 3, and Nikitas Alexandridis

More information

A scheme for racquet sports video analysis with the combination of audio-visual information

A scheme for racquet sports video analysis with the combination of audio-visual information A sheme for raquet sports video analysis with the ombination of audio-visual information Liyuan Xing a*, Qixiang Ye b, Weigang Zhang, Qingming Huang a and Hua Yu a a Graduate Shool of the Chinese Aadamy

More information

Time delay estimation of reverberant meeting speech: on the use of multichannel linear prediction

Time delay estimation of reverberant meeting speech: on the use of multichannel linear prediction University of Wollongong Researh Online Faulty of Informatis - apers (Arhive) Faulty of Engineering and Information Sienes 7 Time delay estimation of reverberant meeting speeh: on the use of multihannel

More information

Gait Based Human Recognition with Various Classifiers Using Exhaustive Angle Calculations in Model Free Approach

Gait Based Human Recognition with Various Classifiers Using Exhaustive Angle Calculations in Model Free Approach Ciruits and Systems, 2016, 7, 1465-1475 Published Online June 2016 in SiRes. http://www.sirp.org/journal/s http://dx.doi.org/10.4236/s.2016.78128 Gait Based Human Reognition with Various Classifiers Using

More information

Machine Vision. Laboratory Exercise Name: Student ID: S

Machine Vision. Laboratory Exercise Name: Student ID: S Mahine Vision 521466S Laoratory Eerise 2011 Name: Student D: General nformation To pass these laoratory works, you should answer all questions (Q.y) with an understandale handwriting either in English

More information

The influence of defeated arguments in defeasible argumentation

The influence of defeated arguments in defeasible argumentation The influene of efeate arguments in efeasible argumentation Bart Verheij University of Limburg, Department of Metajuriia P.O. Box 616, 6200 MD Maastriht, The Netherlans fax: +31 43 256538 email: bart.verheij@metajur.rulimburg.nl

More information

A Novel Bit Level Time Series Representation with Implication of Similarity Search and Clustering

A Novel Bit Level Time Series Representation with Implication of Similarity Search and Clustering A Novel Bit Level Time Series Representation with Impliation of Similarity Searh and lustering hotirat Ratanamahatana, Eamonn Keogh, Anthony J. Bagnall 2, and Stefano Lonardi Dept. of omputer Siene & Engineering,

More information

An Optimized Approach on Applying Genetic Algorithm to Adaptive Cluster Validity Index

An Optimized Approach on Applying Genetic Algorithm to Adaptive Cluster Validity Index IJCSES International Journal of Computer Sienes and Engineering Systems, ol., No.4, Otober 2007 CSES International 2007 ISSN 0973-4406 253 An Optimized Approah on Applying Geneti Algorithm to Adaptive

More information

Torpedo Trajectory Visual Simulation Based on Nonlinear Backstepping Control

Torpedo Trajectory Visual Simulation Based on Nonlinear Backstepping Control orpedo rajetory Visual Simulation Based on Nonlinear Bakstepping Control Peng Hai-jun 1, Li Hui-zhou Chen Ye 1, 1. Depart. of Weaponry Eng, Naval Univ. of Engineering, Wuhan 400, China. Depart. of Aeronautial

More information

Using Augmented Measurements to Improve the Convergence of ICP

Using Augmented Measurements to Improve the Convergence of ICP Using Augmented Measurements to Improve the onvergene of IP Jaopo Serafin, Giorgio Grisetti Dept. of omputer, ontrol and Management Engineering, Sapienza University of Rome, Via Ariosto 25, I-0085, Rome,

More information

On - Line Path Delay Fault Testing of Omega MINs M. Bellos 1, E. Kalligeros 1, D. Nikolos 1,2 & H. T. Vergos 1,2

On - Line Path Delay Fault Testing of Omega MINs M. Bellos 1, E. Kalligeros 1, D. Nikolos 1,2 & H. T. Vergos 1,2 On - Line Path Delay Fault Testing of Omega MINs M. Bellos, E. Kalligeros, D. Nikolos,2 & H. T. Vergos,2 Dept. of Computer Engineering and Informatis 2 Computer Tehnology Institute University of Patras,

More information

Enumerating pseudo-triangulations in the plane

Enumerating pseudo-triangulations in the plane Enumerating pseuo-triangulations in the plane Sergey Bereg Astrat A pseuo-triangle is a simple polygon with exatly three onvex verties. A pseuo-triangulation of a finite point set S in the plane is a partition

More information

Chromaticity-matched Superimposition of Foreground Objects in Different Environments

Chromaticity-matched Superimposition of Foreground Objects in Different Environments FCV216, the 22nd Korea-Japan Joint Workshop on Frontiers of Computer Vision Chromatiity-mathed Superimposition of Foreground Objets in Different Environments Yohei Ogura Graduate Shool of Siene and Tehnology

More information

Spatio-Temporal Naive-Bayes Nearest-Neighbor (ST-NBNN) for Skeleton-Based Action Recognition

Spatio-Temporal Naive-Bayes Nearest-Neighbor (ST-NBNN) for Skeleton-Based Action Recognition Spatio-Temporal Naive-Bayes Nearest-Neighbor () for Skeleton-Based Ation Reognition Junwu Weng Chaoqun Weng Junsong Yuan Shool of Eletrial and Eletroni Engineering Nanyang Tehnologial University, Singapore

More information

Accommodations of QoS DiffServ Over IP and MPLS Networks

Accommodations of QoS DiffServ Over IP and MPLS Networks Aommodations of QoS DiffServ Over IP and MPLS Networks Abdullah AlWehaibi, Anjali Agarwal, Mihael Kadoh and Ahmed ElHakeem Department of Eletrial and Computer Department de Genie Eletrique Engineering

More information