Latent Factor Modeling of Users Subjective Perception for Stereoscopic 3D Video Recommendation
11 Latent Factor Modeling of Users SubjectivePerception for Stereoscopic 3D VideoRecommendation
Balasubramanyam Appina , Member , IEEE , Mansi Sharma , Member , IEEE , Santosh Kumar . Abstract —Numerous stereoscopic 3D movies are released ev-ery year to theaters and created large revenues. Despite theimprovement in stereo capturing and 3D video post-productiontechnology, stereoscopic artifacts which cause viewer discomfortcontinue to appear even in high-budget films. Existing automatic3D video quality measurement tools can detect distortions instereoscopic images or videos, but they fail to consider theviewer’s subjective perception of those artifacts, and how thesedistortions affect their choices. In this paper, we introduce anovel recommendation system for stereoscopic 3D movies basedon a latent factor model that meticulously analyse the viewer’ssubjective ratings and influence of 3D video distortions on theirpreferences. To the best of our knowledge, this is a first-of-its-kind model that recommends 3D movies based on stereo-film-quality ratings accounting correlation between the viewer’s visualdiscomfort and stereoscopic-artifact perception. The proposedmodel is trained and tested on benchmark Nama3ds1-cospad1and LFOVIAS3DPh2 S3D video quality assessment datasets. Theexperiments revealed that resulting matrix-factorization basedrecommendation system is able to generalize considerably betterfor the viewer’s subjective ratings.
Index Terms —Stereoscopic 3D movies, recommender system,matrix factorization, latent factor models, subjective stereo qual-ity assessment, visual discomfort
I. I
NTRODUCTION
The audience of 3D films and virtual reality content isgrowing, as most of the films or YouTube videos have beenreleased in the stereoscopic 3D (S3D) format today. Thereare three popular approaches to generate a stereoscopic 3Dvideo: 1) Scene acquisition using a stereo camera, 2) 2D-to-3D video conversion, which means the creation of left andright eye views from the original source video, 3) Rendering,which is the process of synthesizing views by means of 3Dreconstruction or employing global 3D models and computervision techniques [1]–[5].Despite advances in technology, there are numerous sourcesof visual artifacts to appear in the created stereoscopic pic-ture/video [6], [7]. A comprehensive study of visual artifactsin S3D content has been carried out at MSU Graphics &Media Lab, Moscow State University, under VQMT3D project[6], [7] in cooperation with IITP RAS. The research study The author is with Department of Electronics and Communication Engi-neering, Indian Institute of Information Technology, Design and Manufactur-ing Kancheepuram, Chennai, India, 600127.E-mail: [email protected]. The authors are with Department of Electrical Engineering, Indian Instituteof Technology Madras, Chennai, India, 600036.E-mail: { ee15b110@smail, mansisharma@ee } .iitm.ac.in. identified potential artifacts in several popular HollywoodS3D movies. The artifacts like disparity, scale, color, sharp-ness mismatches or temporal asynchrony, cardboard, crosstalkeffects are prominent in the S3D 3DTV content. Besides,different types of artifacts at various stages of the contentdelivery affect S3D video. The compression, blur and frame-freeze distortions influence 3D video in format-conversion andrepresentation stage, and in the coding and transmission stage[8]. Zeri and Livi [9] interviewed 854 people. They recognizedfrequent symptoms like eye strain, blurred vision and a burningsensation after watching 3D movies in theaters. Even high-budget films, like Pirates of the Caribbean, Dolphin Tale,The Three Musketeers, The Avengers, etc. , contain scenes withgeometric and color impairments, camera rotation difference,shift vertical variation between the left and right views. Animportant research study conducted by Barreda et al. [10]on 3D content using psycho-physiological methods establishcomplex effects of visual discomfort over 3DTV viewer’s emo-tional arousal, which leads to problems like headache, nausea,fatigue and eye strain, etc. The compression artifacts and theirvariation with a depth range of 3D displays noticeably affectsviewer’s perception [11]–[14].The most reliable way to reduce such distortions is tocorrect and enhance the stereoscopic-content quality duringproduction. But the correction process is extremely labor-intensive and heavily relies on the degree of automation and onthe workflow which is not cost efficient. The algorithms for theautomatic detection of such artifacts and quality assessmentare emerging [15], [16]. However, measuring frequency andintensity of an artifact does not account how painful it can befor the viewer. Therefore, it is critical to consider subjectiveperception ratings for artifacts, that is, which egregious distor-tions affect a viewer notably and which distortions are withintolerable limits of his/her visual comfort.In this paper, we proposed a novel recommendation systemfor S3D movies. The well-controlled subjective experimentsand careful statistical analysis conducted by the most studiesestablish that discomfort is greater for some specific distortionsthan for others when viewing stereo video [7], [17]. Mainly theinfluence is from the content itself. We observed most signif-icant information for designing a recommendation system forS3D movies is that describe the viewer’s perceptual discomfortwith the particular distortion types. Despite enough advancesin image/video objective quality assessment techniques [18]–[21], it is difficult to propagate the same achievement for S3Dvideo because automatic estimation of relevant characteristics a r X i v : . [ c s . MM ] J a n for problems that cause visual discomfort is a nontrivialproblem. We wonder when even very simple yet reliablemetrics measure several problems affecting stereo quality onthe fly. Thus, it is crucial to account subjective ratings forhealthy and reasonable 3D video watching as well as properlydesigning of recommendation system.Our recommendation system is built on the latent factormodel that rely on viewer-movie ratings. Given a set ofpristine and distorted S3D videos and their subjective ratings,our latent factor model that is based on matrix factorizationmap viewer’s and 3D videos to a set of latent features. Theproblem of predicting perceptual quality rates of S3D video isformulated as a matrix completion problem for the user-movierating matrix. Our system rate the S3D videos in accordancewith the user’s discomfort level. Our model recommendationmechanism can easily integrate within Netflix matrix factor-ization methods, which is the most important class of col-laborative filtering approaches. The proposed recommendationsystem will be very useful in reducing the flood of low-quality3D content online by ratings stereo 3D more-consistent withquality. The encouraging results obtained by statistical analysisof the proposed model conducted on benchmark Nama3ds1-cospad1 [14] and LFOVIAS3DPh2 [16] S3D data demonstrateits potential for generating accurate predictions.II. R ELATED W ORK
A comprehensive survey of algorithms used by Netflix forits Recommendation System is found in a paper written byLeidy Esperanza MOLINA FERN ´ANDEZ [22]. It coveredCollaborative Filtering, Content-based Filtering, model-basedSVD, PCA, and Probabilistic Matrix Factorization techniques.The paper explains a movie recommendation mechanism buildwithin Netflix on the Matrix Factorization (MF) [23] approachthat learns the latent preferences of users and movies from theratings and make a prediction of the missing ratings estimatingthe dot product of the latent factors [22], [23].Lu et al. [24] applied MF model for computing vectorrepresentations of words. Their work demonstrated how aconvolutional neural network can be integrated into MF modelto produce interpretable recommendations. Lee et al. [25] de-veloped a demo model that considers latest uploaded YouTubevideos. Here the collaborative filtering approach is not muchapplicable since it relies on aggregate user behavior. Instead,they modeled recommendation problem as a video content-based similarity learning. They learned deep video embeddingsand predict ground-truth video relationships from the trainedmodel. However, this approach is built up purely based onvideo content signals.YouTube provides a vast collection of 2D videos. In contrastsince 2009, YouTube offers users an interesting feature toupload two-channel stereo videos for 3D viewing experi-ence. YouTube flash players can support anaglyph videosin red/cyan, green/magenta or blue/yellow layout and followrow/column interlaced display on the screen. The 3D contenton YouTube appears (or display) in accordance with therelevance order. Tsingalis et al. [26] presented a study onYouTube recommendation graphs of 2D and 3D videos. They studied the statistical relevance or recommendation propertiesof social network sites like Facebook, Tweeter and Flickr,such as power-law distribution. Also, they analyzed clusteringmethods to understand the existence of media content groups.Davidson et al. [27] discussed in details about the recommen-dation system in use at YouTube. The study reveals YouTuberecommends personalized sets of videos to users based on theirprevious activity on the web. They discussed some uniquechallenges YouTube faces for video endorsement and howto address them. Covington et al. [28] describes a YouTubesystem at a high level and center their study about substantialperformance improvements brought by deep learning. Theypresented deep architecture built on deep candidate generationsand a separate ranking model for recommending YouTubevideos.Estrada and Simeone [29] developed a recommender systemfor guiding physical object substitution in virtual reality. Thisuser-perception based recommender approach allows themto watch the physical world whilst navigating the virtualenvironment through a video feed. The user identifies thelocation of object placement in the surroundings given thefeedback.Niu et al. [30] presented a video recommendation systembased on the affective analysis of the users. Their subjectivemodel evaluates feature of emotion fluctuation based on theGrey Relational Analysis (GRA). Certain video features areextracted and mapped to the well-known Lovheim emotion-space specifying prominent human feelings, patterns, attitudesand behaviour such as Anger, Distress, Surprise, Fear, Enjoy-ment, Shame, Interest, and Contempt. GRA-based recommen-dation method is developed under the Fisher model to analyseextracted emotions as factors.Zhang et al. [31] developed a recommendation system forMobile AR application incorporating user’s preferences, loca-tion and temporal information in an aggregated random walkalgorithm. Their system predicts user’s preferences modifyingthe graph edge weight and computing the rank score. Similarly,Shi et al. [32] predicts individual location recommendation,Chatzopoulos and Hui [33] anticipates object recommendationin Mobile AR environments.III. M
ATHEMATICAL M ODELING OF
S3D
VIDEOS R ECOMMENDATION S YSTEM
We proposed a novel recommendation system for stereo-scopic 3D videos based on a MF model. In the proposedmodel, viewer’s and S3D movies are mapped to a joint latentfactor space. The row or column associated to a specific vieweror S3D movie is referred as the latent factors. In the mappedlatent factor space of dimensionality, say f , the viewer-movieratings are analyzed as inner products. Suppose each S3Dmovie i is associated with a latent vector q mi ∈ R f , andeach user u is associated with a latent vector p uj ∈ R f . Inthe proposed problem formulation, for a given movie i , theelements of q mi estimate the extent to which the S3D movieholds those factors, whether distorted with a particular artifactor free from that. For a given user u , the elements of p uj determine the extent of user acceptance in S3D movies that (a) Nama3ds1-cospad1 dataset [14]. µ = 3.09, m = 3, σ = 1.35. (b) LFOVIAS3DPh2 dataset [16]. µ = 3.19, m = 3, σ = 1.37. Fig. 1:
Subjective score distribution of dataset. The µ , m and σ denote the statistical measures (mean ( µ ), median ( m ), standard deviation( σ )) of the subjective ratings. are high on the corresponding factors, again, whether distortedwith a particular artifact or not. The model approximatesviewer u (cid:48) s rating of S3D movie i by measuring resultingdot product, ˆ r ui = q m T i p uj . The dot product captures in-terconnection between the viewer u and S3D movie i , thatis, the viewer’s overall acceptance/tolerance in the particulardistortion viewing the movies. Once the mapping is computedfor each S3D movie and viewer to factor vectors q mi , p uj ∈ R f ,the proposed model easily determines the rating a viewer willgive to any S3D movie with distortions by using ˆ r ui .We avoided imputation in proposed model [34]. The ob-served ratings are modeled directly as suggested by [23],[35] and avoided overfitting through the regularization. On theset of known matrix ratings, the regularized squared error isminimized to learn the factor vectors q mi , p uj as min ˆ p, ˆ q (cid:88) ( u,i ) ∈ S ( r ui − q m T i p uj ) + λ ( || q mi || + || p uj || ) (1)where, S is the training set of ( u, i ) pairs for which r ui isknown.To make matrix factorization approach more effective in ourproposed application-specific requirements, we add biases incapturing the full ratings of the observed signals ˆ r ui = µ + b i + b u + q m T i p uj (2)The observed rating in (2) is broken down into its four compo-nents: global average (or mean), 3D movie bias, viewer bias,and viewer-movie interaction. This allows each component torepresent only the part of an observed signal relevant to it. Themodel is learned by minimizing the squared error function as min ˆ p, ˆ q, ˆ b (cid:88) ( u,i ) ∈ S ( r ui − µ − b i − b u − q m T i p uj ) + λ ( || q mi || + || p uj || + b i + b u ) (3) The stochastic gradient descent algorithm [23], [36], [37] isused to optimize equation (3). For better accuracy in predic-tion, the algorithm loops through all ratings in the training dataand estimate the model parameters. The system estimates ˆ r ui for each given training case. The prediction error is determinedas E ui = r ui − µ − b i − b u − q m T i p uj (4)The parameters are updated as b i ← b i + ς ( E ui − λb i ) (5) b u ← b u + ς ( E ui − λb u ) (6) q mi ← q mi + ρ ( E ui p uj − λq mi ) (7) p uj ← p uj + ρ ( E ui q mi − λp uj ) (8)where, ρ and ς specify constant magnitudes that accountsproportion by which parameters are modified in the oppositedirection of the gradient.The objective of our matrix factorization model is to the pre-dicts the unknown future S3D video ratings, from the learnedmodel obtained by fitting the earlier observed ratings. Wedetermined the regularization constant λ by cross-validation[38]. IV. R ESULTS AND D ISCUSSION
The efficacy of the proposed algorithm is evaluated on theNama3ds1-cospad1 [14] and LFOVIAS3DPh2 [16] S3D videodatasets. Nama3ds1-cospad1 database has 10 reference and100 test S3D video sequences. The video sequences have aresolution of × and saved in .avi container. Theframe rate is 25 fps and a duration of either 16 sec or 13 sec.The database is a combination of H.264 and JP2K, scaling anddown sampling distorted S3D video sequences. These artifactsare applied symmetrically on each view of an S3D videoand published the Difference Mean Opinion Score (DMOS)scores as subjective scores. Human assessment on perceptual (a) Pristine stereoscopic video frame. (b) Distorted stereoscopic video frame.(c) Pristine stereoscopic video. (d) Distorted stereoscopic video. Fig. 2:
Perceptual opinion score of each subject and proposed algorithm prediction on pristine and distorted ‘Hall’ S3D videos from theNama3ds1-cospad1 [14] dataset. quality was performed in single stimulus continuous qualityevaluation (SSCQE) with hidden reference method. They haveused 5 scales to rate the perceptual quality of an S3D videoand 28 subjects involved in the study. They have publishedeach subject quality score and an overall mean quality scoreof the dataset. The LFOVIAS3DPh2 S3D video dataset has12 pristine sequences with good variety of structure, texture,depth and temporal information. The video sequences have aresolution of × and duration of 10 seconds witha frame speed of 25 fps. They created 288 test stimuli byintroducing the H.264 and H.265 compression, blur and framefreeze distortions. The dataset is a combination of symmetricand asymmetric S3D videos. They have used SSCQE methodto perform the subjective study and 20 subjects involved inthe study. They published each subject perceptual qualityscore and final DMOS of the dataset. Figure 1 shows theviewer perceptual score distribution of Nama3ds1-cospad1 andLFOVIAS3DPh2 S3D video datasets. From the plot, it is clearthat both datasets are diverse in perceptual video quality range.Also, the estimated statistics ( µ , m and σ ) of each dataset areconsistent and followed the observed trend in the perceptualquality of S3D videos.Figure 2a shows the st frame from left view of the pristine‘Hall’ S3D video from Nama3ds1-cospad1 dataset. Figure2b shows the st frame of H.264 (Quantization Parameter= 38) compressed S3D video of the corresponding referenceview. Figures 2c and 2d show the distribution of subjective assessment rates and proposed algorithm predicted perceptualquality rates of a pristine and distorted S3D video, respectively.From the plot it is clear that the proposed algorithm accuratelypredicts the subjective quality rates of pristine and distortedvideos. Also, the deviation between average scores of subjec-tive rates and the proposed algorithm predictions is very less.The plot clearly demonstrates the proposed algorithm efficacyto model the perceptual subjective quality ratings of a givenS3D video.The performance of the proposed algorithm is measuredusing the Linear Correlation Coefficient (LCC), SpearmanRank Order Correlation Coefficient (SROCC) and Root MeanSquare Error (RMSE). LCC indicates the linear dependencebetween two quantities. The SROCC measures monotonicrelationship between two input sets. RMSE measures the mag-nitude error between estimated scores and subjective scores.Higher LCC and SROCC values indicate good agreementbetween subjective and objective measures, and lower RMSEsignifies more accurate prediction performance. For both thedatabases, 80% of the human opinion scores is used forproposed algorithm training and the remaining samples areused for testing. In other words, the non-overlapped trainingand test sets are obtained by partitioning the set of availablehuman opinion scores in the 80:20 proportion. We performedthe random assignment for 100 trials of each epoch forstatistical consistency and repeated it for 200 epochs. Finally,we calculated the mean of the LCC, SROCC and RMSE TABLE I:
Performance evaluation of proposed algorithm on Nama3ds1-cospad1 [14] and LFOVIAS3DPh2 [16] video dataset subjectivescores.
Score Training Set Testing SetLCC SROCC RMSE LCC SROCC RMSE
Nama3ds1-cospad1 [14] 0.8873 0.8858 0.6903 0.8753 0.8700 0.7527LFOVIAS3DPh2 [16] 0.8966 0.8911 0.5809 0.8585 0.8288 0.6522
TABLE II:
Proposed algorithm performance on each distortion typeof LFOVIAS3DPh2 S3D video dataset [16].
Type H.264 H.265 Blur FF OverallLCC 0.910 0.904 0.877 0.908 0.858SROCC 0.902 0.914 0.828 0.899 0.828RMSE 0.357 0.384 0.423 0.371 0.652 measures of all epochs to report the performance analysis.Table I shows the performance evaluation of the proposedalgorithm on the training and test sets of Nama3ds1-cospad1and LFOVIAS3DPh2 S3D video datasets. It is clear that theproposed algorithm shows robust performance across both thedatasets.Fig. 3: Variation of proposed algorithm LCC score over 100trails of an epoch. Standard deviation of LCC score over 100trails is × − .Figure 3 shows the LCC score variation of 100 iterationsof an epoch. From the plot it is clear that the scores areconsistent across all iterations, and further, we experiencedthe lower standard deviation (2 × − ) of 100 LCC scores.Figure 4 shows the average training and test RMSE measurevariation over 200 epochs. From the plot, it is clear that boththe RMSE errors reduced with the number of epochs. Theseplots clearly demonstrate the proposed algorithm efficacy toestimate the human assessment quality of a given video. TableII shows the proposed method performance on each distortiontype of LFOVIAS3DPh2 S3D video dataset. The proposedmethod clearly demonstrates the consistent performance acrossall distortion types of the LFOVIAS3DPh2 S3D video dataset.V. C ONCLUSION
This paper presented a novel recommendation system forS3D movies. This is a first attempt that accounts 3DTV Fig. 4: RMSE error variation across epochs.viewer’s subjective ratings for visual artifacts and analysetheir degree of visual discomfort to predict “rating” or “pref-erence” that the viewer’s would give to the S3D movie.In this study, we considered common spatial and temporaldistortion types; JPEG, Resolution reduction, Downsampling& sharpening, Image sharpening, Blur, Frame-freeze, H.264and H.265 compressions; that adversely affect S3D videosignal at different stages of the content generation and deliverychain. Experimental results on 3DTV viewer’s subjective studyand parameter evaluation of latent factors demonstrate that theproposed matrix factorization based model improve accuracyof S3D video affective analysis and performance of recom-mendation. This model will be very useful for media-serviceproviders like Netflix, Amazon, TiVo to recommend quality3D videos and minimize flood of low-quality content basedon the viewer’s subjective perception, depending on their agegroups and preferences.We will further extend this recommendation system byconsidering the detail analysis of commercial S3D movies.The model will be improved by offering per-frame analysisof artifacts causing potential visual discomfort while viewingstereo films like large horizontal disparity, vertical parallax,crosstalk noticeability, cardboard effect, stuck-to-backgroundobjects, stereo window violation, depth continuity, etc. Suchartifacts earn poor ratings according to the existing metrics.Combining objective and subjective scores will help reducethe error rate further while recommending new stereo movies.Besides, we will perform affective analysis on the emotionalreactions of 3DTV viewers while watching stereo 3D moviesor virtual reality S3D content. We will account both subjectivescores and brain-activity measurements to understand the dependencies between the degree of viewer discomfort and theintensity of the distortions. This will help to better classifyviewer’s from different age groups by their susceptibility toartifacts and movies content types. How this affect viewer’saccumulation of discomfort caused by stereoscopic movies andinfluence recommendation ratings is an interesting endeav-our of future study ?. In future work, we will account thepercentage of viewers susceptible to various distortions. Wewill design new experiments and work on evaluation modelslike probabilistic matrix factorization (PMF) to improve thepredictive accuracy. We will experiment on the linear combi-nation of predictions of multiple PMF models with predictionsof Restricted Boltzmann Machine (RBM) models. This couldsignificantly improve the accuracy of the blended solution.R
EFERENCES[1] A. Smolic, P. Kauff, S. Knorr, A. Hornung, M. Kunter, M. Muller,and M. Lang, “Three-dimensional video postproduction and processing,”
Proceedings of the IEEE , vol. 99, no. 4, pp. 607–625, 2011.[2] M. Sharma, S. Chaudhury, B. Lall, and M. Venkatesh, “A flexiblearchitecture for multi-view 3dtv based on uncalibrated cameras,”
Journalof Visual Communication and Image Representation , vol. 25, no. 4,pp. 599–621, 2014.[3] M. Sharma,
Uncalibrated camera based content generation for 3D multi-view displays . PhD thesis, 2017.[4] M. Sharma, S. Chaudhury, and B. Lall, “A novel hybrid kinect-variety-based high-quality multiview rendering scheme for glass-free 3d dis-plays,”
IEEE Transactions on Circuits and Systems for Video Technology ,vol. 27, no. 10, pp. 2098–2117, 2016.[5] E. Cheng, P. Burton, J. Burton, A. Joseski, and I. Burnett, “Rmit3dv:Pre-announcement of a creative commons uncompressed HD 3D videodatabase,” in
Fourth International Workshop on Quality of MultimediaExperience , pp. 212–217, July 2012.[6] A. Antsiferova and D. Vatolin, “The influence of 3d video artifacts ondiscomfort of 302 viewers,” in , pp. 1–8, IEEE, 2017.[7] “Msu 3d-video quality analysis, video quality measurementtool 3d project, msu graphics & media lab (video group),http://compression.ru/video/vqmt3d/.”[8] A. Gotchev, G. B. Akar, T. Capin, D. Strohmeier, and A. Boev, “Three-dimensional media for mobile devices,”
Proceedings of the IEEE ,vol. 99, no. 4, pp. 708–741, 2011.[9] F. Zeri and S. Livi, “Visual discomfort while watching stereoscopicthree-dimensional movies at the cinema,” ophthalmic and PhysiologicalOptics , vol. 35, no. 3, pp. 271–282, 2015.[10] M. Barreda-Angeles, R. P´epion, E. Bosc, P. Le Callet, and A. Pereda-Banos, “How visual discomfort affects 3dtv viewers’ emotional arousal,”in , pp. 1–4, IEEE, 2014.[11] B. Appina, K. Manasa, and S. S. Channappayya, “Subjective andobjective study of the relation between 3d and 2d views based on depthand bitrate,”
Electronic Imaging , vol. 2017, no. 5, pp. 145–150, 2017.[12] S. Jumisko-Pyykk¨o, T. Haustola, A. Boev, and A. Gotchev, “Subjectiveevaluation of mobile 3d video content: depth range versus compressionartifacts,” in
Multimedia on Mobile Devices 2011; and MultimediaContent Access: Algorithms and Systems V , vol. 7881, p. 78810C,International Society for Optics and Photonics, 2011.[13] G. Sanchez, J. Silveira, L. V. Agostini, and C. Marcon, “Performanceanalysis of depth intra-coding in 3d-hevc,”
IEEE Transactions on Cir-cuits and Systems for Video Technology , vol. 29, no. 8, pp. 2509–2520,2018.[14] M. Urvoy, M. Barkowsky, R. Cousseau, Y. Koudota, V. Ricorde,P. Le Callet, J. Gutierrez, and N. Garcia, “Nama3ds1-cospad1: Subjec-tive video quality assessment database on coding conditions introducingfreely available high quality 3d stereoscopic sequences,” in , pp. 109–114, IEEE, 2012.[15] A. Bokov, D. Vatolin, A. Zachesov, A. Belous, and M. Erofeev, “Au-tomatic detection of artifacts in converted s3d video,” in
StereoscopicDisplays and Applications XXV , vol. 9011, p. 901112, InternationalSociety for Optics and Photonics, 2014. [16] B. Appina, S. V. R. Dendi, K. Manasa, S. S. Channappayya, and A. C.Bovik, “Study of subjective quality and objective blind quality predic-tion of stereoscopic videos,”
IEEE Transactions on Image Processing ,vol. 28, no. 10, pp. 5027–5040, 2019.[17] J. Li, M. Barkowsky, and P. Le Callet, “Visual discomfort in 3dtv:Definitions, causes, measurement, and modeling,” in
Novel 3D MediaTechnologies , pp. 185–209, Springer, 2015.[18] F. Speranza, W. J. Tam, R. Renaud, and N. Hur, “Effect of disparityand motion on visual comfort of stereoscopic images,” in
ElectronicImaging 2006 , pp. 60550B–60550B, International Society for Opticsand Photonics, 2006.[19] M.-J. Chen, C.-C. Su, D.-K. Kwon, L. K. Cormack, and A. C. Bovik,“Full-reference quality assessment of stereopairs accounting for rivalry,”
Signal Processing: Image Communication , vol. 28, no. 9, pp. 1143–1155, 2013.[20] B. Appina, M. K., and S. S. Channappayya, “A full reference stereo-scopic video quality assessment metric,” in
International Conference onAcoustics, Speech and Signal Processing, IEEE , pp. 2012–2016, March2017.[21] B. Appina and S. Channappayya, “Full-reference 3-D video quality as-sessment using scene component statistical dependencies,”
IEEE SignalProcessing Letters , vol. 25, pp. 823–827, June 2018.[22] L. E. M. FERN ´ANDEZ, “Recommendation system for netflix,”
Facultyof Science Business Analytics, Vrije Universiteit Amsterdam , 2018.[23] Y. Koren, R. Bell, and C. Volinsky, “Matrix factorization techniques forrecommender systems,”
Computer , vol. 42, no. 8, pp. 30–37, 2009.[24] Y. Lu, R. Dong, and B. Smyth, “Convolutional matrix factorization forrecommendation explanation,” in
Proceedings of the 23rd InternationalConference on Intelligent User Interfaces Companion , pp. 1–2, 2018.[25] J. Lee, N. Kothari, and P. Natsev, “Content-based related video rec-ommendations,”
Advances in Neural Information Processing Systems(NIPS) Demonstration Track , 2016.[26] I. Tsingalis, I. Pipilis, and I. Pitas, “A statistical and clustering study onyoutube 2d and 3d video recommendation graph,” in , pp. 294–297, IEEE, 2014.[27] J. Davidson, B. Liebald, J. Liu, P. Nandy, T. Van Vleet, U. Gargi,S. Gupta, Y. He, M. Lambert, B. Livingston, et al. , “The youtube videorecommendation system,” in
Proceedings of the fourth ACM conferenceon Recommender systems , pp. 293–296, 2010.[28] P. Covington, J. Adams, and E. Sargin, “Deep neural networks foryoutube recommendations,” in
Proceedings of the 10th ACM conferenceon recommender systems , pp. 191–198, 2016.[29] J. G. Estrada and A. L. Simeone, “Recommender system for physicalobject substitution in vr,” in , pp. 359–360, IEEE, 2017.[30] J. Niu, S. Wang, Y. Su, and S. Guo, “Temporal factor-aware videoaffective analysis and recommendation for cyber-based social media,”
IEEE Transactions on Emerging Topics in Computing , vol. 5, no. 3,pp. 412–424, 2017.[31] Z. Zhang, S. Shang, S. R. Kulkarni, and P. Hui, “Improving augmentedreality using recommender systems,” in
Proceedings of the 7th ACMconference on Recommender systems , pp. 173–176, 2013.[32] Z. Shi, H. Wang, W. Wei, X. Zheng, M. Zhao, and J. Zhao, “A novelindividual location recommendation system based on mobile augmentedreality,” in , pp. 215–218, IEEE,2015.[33] D. Chatzopoulos and P. Hui, “Readme: A real-time recommendationsystem for mobile augmented reality ecosystems,” in
Proceedings of the24th ACM international conference on Multimedia , pp. 312–316, 2016.[34] D. Bertsimas, C. Pawlowski, and Y. D. Zhuo, “From predictive methodsto missing data imputation: an optimization approach,”
The Journal ofMachine Learning Research , vol. 18, no. 1, pp. 7133–7171, 2017.[35] A. Paterek, “Improving regularized singular value decomposition forcollaborative filtering,” in
Proceedings of KDD cup and workshop ,vol. 2007, pp. 5–8, 2007.[36] Y. Koren, “Factorization meets the neighborhood: a multifaceted col-laborative filtering model,” in
Proceedings of the 14th ACM SIGKDDinternational conference on Knowledge discovery and data mining ,pp. 426–434, 2008.[37] B. Jin and X. Lu, “On the regularizing property of stochastic gradientdescent,”
Inverse Problems , vol. 35, no. 1, p. 015004, 2018.[38] A. Mnih and R. R. Salakhutdinov, “Probabilistic matrix factorization,”in