Suitability Analysis of Holographic vs Light Field and 2D Displays for Subjective Quality Assessment of Fourier Holograms
Ayyoub Ahar, Maksymilian Chlipala, Tobias Birnbaum, Weronika Zaperty, Athanasia Symeonidou, Tomasz Kozacki, Malgorzata Kujawinska, Peter Schelkens
SSUBMITTED TO THE JOURNAL OF IEEE TRANSACTIONS ON MULTIMEDIA,20 SEPTEMBER 2019 1
Suitability Analysis of Holographic vs Light Fieldand 2D Displays for Subjective Quality Assessmentof Fourier Holograms
Ayyoub Ahar,
Member, IEEE,
Maksymilian Chlipala, Tobias Birnbaum,
Member, IEEE,
Weronika Zaperty,Athanasia Symeonidou,
Member, IEEE,
Tomasz Kozacki, Malgorzata Kujawinska,and Peter Schelkens,
Member, IEEE
Abstract —Visual quality assessment of digital holograms isfacing many challenges. Main difficulties are related to thelimited spatial resolution and angular field of view of holographicdisplays in combination with the complexity of steering andoperating them for such tasks. Alternatively, non-holographicdisplays – and in particular light-field displays – can be utilizedto visualize the numerically reconstructed content of a digitalhologram. However, their suitability as alternative for holo-graphic displays has not been validated. In this research, we haveinvestigated this issue via a set of comprehensive experiments.We used Fourier holographic principle to acquire a diverse setof holograms, which were either computer-generated from pointclouds or optically recorded from real macroscopic objects. Afinal public data set comprising 96 holograms was created usingthree compression methods which encoded the holograms atfour bit-depths. Three separate subjective-tests were conductedusing a holographic display, a light field display and a 2Ddisplay. For these subjective experiments, a double stimulus,multi-perspective, multi-depth subjective testing methodologywas designed and implemented. The tests show that the non-holographic displays indicate a higher sensitivity to artifactsthan the holographic display, though at the same time it isdemonstrated they are highly correlated. This indicates that thenumerically reconstructed holograms rendered on a light field or2D display have a high predictive value for the perceived qualityon holographic display.
Index Terms —Quality Assessment, Holography, SubjectiveTest, Fourier Holography, Holographic Display, Perceptual Qual-ity, Light-Field.
I. I
NTRODUCTION D IGITAL holography, in theory, is considered to be theholy grail of 3D imaging solutions [1]. While the conceptand its initial realizations has been around for almost half acentury, only recently it is gaining again interest for 3D visu-alization. This is due to steady growth of available computa-tional power and significant improvements in nano-electronics,optical hardware and photonics technologies. However, quitea few hardware and signal processing challenges are yet to beaddressed in order to facilitate an immersive 3D experience
A. Ahar, T. Birnbaum, A. Symeonidou and P. Schelkens are with theDept. of Electronics and Informatics (ETRO), Vrije Universiteit Brussel(VUB),Pleinlaan 2, B-1050 Brussels, Belgium and imec, Kapeldreef 75, B-3001 Leuven, Belgium e-mail: [email protected]. Chlipala, W. Zaperty, T. Kozacki and M. Kujawinska are with WarsawUniversity of Technology, Institute of Micromechanics and Photonics, 8 Sw.A. Boboli St., 02-525 Warsaw, Poland. via a complete pipeline for high-quality dynamic holographywith full-parallax and wide field of view (FoV) [2].In this regard, one of the core challenges is modeling theperceived visual quality of the rendered holograms, whichhas a vital impact on steering the other components of theholographic imaging pipeline. While the design of highlyefficient numerical methods in Computer-Generated Holog-raphy (CGH) [3], [4], [5], [6], [7] and efficient encodersfor holographic content [2], [8], [9], [10], [11] is gainingmomentum, Visual Quality Assessment (VQA) of hologramshas a rather long way to reach its primary milestones dueto various open problems along the way [2], [11]. Indeed,conducting a systematic subjective test and creating a scoreddatabase from a diverse set of holograms is the very firststep. But this by itself reveals to be a challenging task. Notonly there is no widely-accepted standard methodology forplenoptic content and especially for holographic data, but alsoholographic displays with acceptable visual characteristics arestill scarce. Moreover, configuring and operating such displaysrequires advanced technical skills. Often, researchers havebeen rendering numerically reconstructed holograms on non-holographic displays, including regular 2D displays or morerecently multi-view light-field displays, to alleviate this prob-lem [12], [13], [14]. However, potential perceptual differencesbetween visualization on holographic displays and numericalreconstructions rendered on non-holographic displays to ourknowledge have not been investigated thoroughly before.Some of the most evident issues include loss of visual cuesrelated to the depth perception on 2D displays and light-fielddisplays, FoV and appearance of different types of specklenoise.These issues are inter-connected with the chosen displayfor visualizing the holographic content. For example for a2D display, only a specific focus plane and perspective ofthe hologram can be rendered. For a light-field display, foreach view the hologram needs to be reconstructed for aparticular focus plane utilizing a suitable aperture. As such,only a section of the 3D scene volume described by thehologram is rendered properly. Nevertheless, both displayscan support high spatial resolution and large display sizes. Onthe other hand, holographic displays can render the completeplenoptic scene, but their resolution and overall size arecurrently limited, which in practice results in only a tiny view-ing window(VW) to explore the visualized hologram. These a r X i v : . [ ee ss . I V ] S e p UBMITTED TO THE JOURNAL OF IEEE TRANSACTIONS ON MULTIMEDIA,20 SEPTEMBER 2019 2 fundamentally diverse properties require different strategiesand procedures per display type to conduct a subjective test.The main objectives and novelties of this manuscript in-clude:1) Comparing holographic versus non-holographic displaysbased on the visual appearance of same set of holo-grams;2) Designing and implementing a test methodology forsubjective testing of holograms;3) Creating the first publicly available database of opticallyrecorded and computer-generated holograms annotatedwith subjective test results;4) Evaluation of computer-generated against opticallyrecorded Fourier holograms.In section II, the holographic, light field and regular 2Ddisplay are described that are being compared in this test toassess their suitability to evaluate the visual quality of holo-grams. The details about the numerical and optical methods toproduce the holograms used in this experiment are providedin section III as well as the content preparation. Section IVintroduces the subjective test methodology including its detailsfor each setup and training of the test subjects. In section V, weexplain the statistical post-processing of the experimental re-sults and provide the analysis and discussion of the outcomes.Finally, section VI presents the concluding remarks.II. D
ISPLAY SYSTEMS
A. Holographic display
In this work, a Fourier holographic display with an inco-herent LED source is employed. The system provides high-quality orthoscopic reconstructions of large objects [15], whichcan be viewed with a naked eye. Also, it facilitates a stableperformance through very deep scenes [16]. The display setupis presented in the Fig. 1. In this system, a phase-only spatiallight modulator (SLM) (Holoeye 1080P, 1920 × µ m) is illuminated by a normal plane wave,which is formed by an LED source (Doric Lenses, centerwavelength λ G = 515 nm and fiber core of 960 µ m) and acollimating lens L C (F C = 400 mm). The SLM is conFigd todisplay the object wave with removed spherical phase factor.Next, the reflected beam passes through the imaging module,which introduces a magnification and facilitates the complexwave coding. The first imaging element is realized by a 4Fafocal imaging system composed of the lenses L (F = 100mm) and L (F = 600 mm) with magnification ratio M = -6.The 4F system and the field lens L f conjugate the SLM planewith a 3D hologram reconstruction volume focused on theVW. The complex coding scheme is experimentally supportedwith the absorbing cut-off filter in the Fourier plane of the 4Fsystem [17].In this experiment, all the optical components on the opticaltable were covered using black colored barriers such that noenvironmental light would enter the black box, i.e. the displaysetup. A small slit was carved into the box and a metal chinrestand forehead holder were put in front of the slit such that allsubjects could easily observe the displayed holograms as soonas they would position their head accordingly. Figure 1: Fourier holographic display setup.The Fourier holography enables reconstruction of a 1:1orthoscopic copy of the 3D object with no visible distortionsfor the holographic display [18] described above. The fullobject is viewed by the naked eye and objects with a maximumsize of 107 mm can be observed from a distance of 700mm. With the available Space Bandwidth Product (SBP) [19],[20] of the SLM this results in an angular FoV = 8.8 o andan angular resolution of display which is comparable to theresolution of the human eye for dark observation conditions.The 2D and light field displays discussed below are based on2D reconstructions of a single or multiple views, respectively.Imaging on both displays benefits from the convention ofFourier holography as well since it provides full use of theSBP and thereby achieves the highest quality during therecording/generation process. B. 2D display
The issued 2D display is a professional Eizo CG318-4Kmonitor with 4K UHD resolution (3840 × m brightness, and minimum black level of 0.2cd/ m . On this display, numerical reconstructions in the objectplane were rendered for a particular reconstruction distanceand perspective. C. Light field display
The light field display system is a HoloVizio-722RC byHolografika [22]. This is a 72 inch display having an horizontalangular FoV of 70 o with a total 3D resolution of 73 Mpixel.It provides a 2D equivalent resolution of 1280 ×
768 pixels foreach of the 72 views. It provides a 24-bit RGB color systemwith a brightness of ≈ m . Holograms are rendered onthis display by calculating numerically the reconstructions fora particular reconstruction distance for each view supportedby the display. III. T EST D ATA
For a successful subjective experiment the test data needto provide sufficient diversity in terms of the features ofthe represented 3D scenes, production of the holograms and
UBMITTED TO THE JOURNAL OF IEEE TRANSACTIONS ON MULTIMEDIA,20 SEPTEMBER 2019 3
Figure 2: An abstract schematic of the experimental pipeline.distortions introduced. Having a set of holograms obtainedfrom a diverse set of objects is vital to avoid any bias in theresults. This is particularly important in holography since eachhologram is an interferogram. Therefore, characteristics of therecorded scene (e.g. object positions, their distances to therecording plane, surface properties and occlusions) will affectthe entire interference footprint on the recorded hologram.On top of the scene characteristics, the method whichproduces the hologram must be taken into account. Hologramscan either be optically recorded or numerically computed.The latter category, CGHs, can nowadays be calculated atvery high spatial resolutions allowing support for high SBPs,efficient occlusion handling and Bidirectional Reflectance Dis-tribution Functions (BRDFs). However, photo-realistic qualityis difficult to achieve. Therefore, it is also important toinclude Optically Recorded Holograms (ORHs) in the testdata. Moreover, ORHs have different characteristics such asthe presence of incoherent measurement noise.Four ORHs from objects of various dimensions, surfacecharacteristics and capture distances, were selected, which areshown in Fig. 2. The first object is a “Mermaid” figurinewith small depth, while the second object is a “Squirrel”figurine with larger depth. Both objects are characterized bya glossy, metallic surface. The third and the fourth objects,“Wolf”, a rubber toy, and “Sphere”, the 3D printed modelbased on the input content of the CGH "Ball", respectivelyhave diffuse surfaces and rather large depths. The hologramsof these real objects were recorded using a lensless Fourierholographic capture system [23], described in section III-A.The holograms of the synthetic objects, represented as pointclouds, are generated with a multiple Wavefront RecordingPlane (WRP) method [3] shortly discussed in section III-B.All considered holograms are created with respect to aspherical reference wave with focal point in the scene center.In the Fourier holographic capture system a spherical referencepoint source is placed at the center of the object plane. In the CGH calculation framework a demodulation with a Fresnelapproximated spherical phase factor is performed numericallyafter initial propagation to the hologram plane. In this way theadvantage of the lensless Fourier holographic capture systemin terms of SBP [24] is utilized in both scenarios. This meansthat the hologram pixel count does not limit the maximalobject dimensions but instead the maximal FoV. ORHs andCGHs are obtained for a high resolution of 16384 × A. Optical acquisition
For optical acquisition a lensless, Fourier synthetic apertureholographic capture system [23], [25] is employed (Fig. 3).The laser beam is divided into a reference and an object beamby a polarizing beam splitting cube PBS. The intensity ratioof both beams is adjusted with an achromatic half-wave plate λ /2, to obtain a high contrast for the interference fringes ofa given scene. The reference beam is formed and directed bythe following set of elements: a pinhole PH, an achromaticcollimating lens C (F C = 300 mm, NA C = 0.13), and mirrorsM and M . The reference point source S is generated at theobject plane by an achromatic objective L(F L = 60 mm, NA L = 0.21). The lenses C and L are selected such that they coverthe entire area of the synthetic aperture hologram capture. Thediffusers D and D create a double-sided illumination withthe help of the mirrors M , M , M and another beam splittingcube BS. The analyzer A, placed in front of the camera,improves the hologram contrast by filtering out non-interferinglight. The hologram is recorded by a charge-coupled device(CCD) camera (Basler piA2400-12gm) with a pixel pitch of3.45 µ m and a resolution of 2448 × between adjacent captured sub-holograms and enables UBMITTED TO THE JOURNAL OF IEEE TRANSACTIONS ON MULTIMEDIA,20 SEPTEMBER 2019 4
Figure 3: Lensless Fourier synthetic aperture holographiccapture system.data stitching with sub-pixel precision using a correlation-based routine [26]. The obtained off-axis, synthetic aperturelensless Fourier holograms are composed of 20 sub-hologramsthat have a physical size of approximately 56.5 mm × × m adapted to thescene size and with either 532 nm or 632.8 nm laser beamwavelengths λ n : “Mermaid”, R = 450 mm, λ = 532 nm;“Squirrel”, R = 500 mm, λ = 632.8 nm; “Wolf”, R = 780mm, λ = 532 nm; and “Sphere”, R = 960 mm with λ =532 nm. B. Computer-generated holograms
The CGHs, used for the subjective tests, were generatedfrom point clouds with an extension of the WRP method [3].This method employs multiple parallel wavefront recordingplanes and pre-computed look-up tables. Moreover, it includesan occlusion handling technique. As shown in Fig. 4, thecontribution of each point – starting from the point furthestaway to the hologram plane to the closest point – is addedto the respectively closest WRP. When all the points thatbelong to the current WRP are accounted for the wavefieldis propagated to the next WRP and so forth. To simulatediffuse reflection we assign a random phase to the pointspread function during the calculation of the LUTs, as pre-sented in [27]. However, there is a very important difference,compared to the previously published methods. To exploitthe SBP advantage of the Fourier holographic approach, thewavefield at the last WRP plane is converted to comply withthe Fourier hologram configuration, contrary to the in-planeconfiguration that it supported before. This is done in twosteps. First, hologram is propagated to its proper viewingdistance using the angular spectrum method[28], [29] andsubsequently demodulated with a quadratic Fresnel phase ker-nel corresponding to the axial distance between the hologramand the last WRP. The second step approximates a sphericalwavefront with focus in the center-plane of the object by usingthe Fresnel approximation.The four CGHs have the same setup parameters: the pixelpitch is 3.45 µ m, the wavelength of the reference beam is 532 nm and the scene center plane were located 700 mm from thehologram plane.Figure 4: Illustration of the multiple-WRP CGH method usedfor the generation of the CGHs [3]. Additionally, the variationof the support of the PSF per depth level of the LUT is shown,which is determined by the distance to the WRP and themaximum diffraction angle. C. Content preparation
Finally, to facilitate subjective testing and to examine thesuitability of each used display, the produced holograms haveto be processed such that they are available at different visualquality levels. This enables testing the sensitivity of eachdisplay for quality degradation of the holographic content. Anadded complexity here is the selection of suitable distortiontypes. Nonetheless, compression artifacts are a good startingpoint for a holographic dataset, both from a practical point ofview, comparing the performance of the available compressionmethods, and also based on the fact that their artifacts normallystem from a combination of multiple distortion types as aresult of different processes undergone inside the encoders.Though all classically used distortions in visual quality testingcould be considered, it is important to realize that the end-user will observe the reconstructed hologram in the objectplane and not in the hologram plane. During the reconstructionor back-propagation process the propagated data from eachpoint on the fringe pattern updates each and every point ofthe reconstructed scene. Hence, the reconstructed scene isparticularly resilient to local artifacts or even complete lossof information in some small regions of the hologram. Asan example, salt and pepper noise, which in regular imagingsignificantly degrades the visual quality, almost completelyvanishes after reconstruction of the hologram. Therefore, inthis experiment we constrained the distortions to those thathave a more global impact on the hologram, more particularlycompression distortions. We employed three coding engines:JPEG 2000 [30], [31], intra H.265/HEVC [32] and wave atomcoding (WAC) [33].The IRIS-JP3D software package was deployed to imple-ment the JPEG 2000 compression [34], [35]. The defaultconfiguration for JPEG 2000 was utilized using a 4-levelMallat decomposition and CDF 9/7 wavelets with 64 × UBMITTED TO THE JOURNAL OF IEEE TRANSACTIONS ON MULTIMEDIA,20 SEPTEMBER 2019 5
Table I: Characteristics of the objects utilized to generate the holograms.
Hologram Aquisition Method PC density/Material No. WRP Recording Dist(mm) Obj. Size: W × H × D(mm) Rec.Dist/DepthOR-Mermaid
ORH Polished Metal - 450 27 × × OR-Ball
ORH 3D-Print - 960 65 × ×
65 14.76
OR-Squirrel
ORH Brushed Metal - 500 43 × ×
70 7.14
OR-Wolf
ORH Plastic - 780 50 × ×
80 9.12
CG-Ball
CGH 1.313.280 101 700 50 × ×
50 14.00
CG-Chess
CGH 219.100 200 491 38 × ×
310 1.58
CG-Earth
CGH 306.372 101 706 46 × ×
46 15.34
CG-Plane
CGH 9.999.079 200 716 53 × ×
71 10.08
Figure 5: Center-views of numerical reconstructions for the reference holograms generated and utilized for this experiment.The top row contains the 4 CGHs from point-clouds and the bottom row shows the ORHs from real objects. The "Sphere"hologram was recorded from the 3D print of the "Ball" Point-Cloud.For the experiments, revision HM-16.18 [36] was used asimplementation of the H.265/HEVC compression standard.Since all tested images were in grayscale format, we used 4:0:0subsampling and fed the images as an 8-bit luminance channelwith empty chrominance channels. Hence, cross-componentprediction and motion search settings were disabled. "Framerate" and "frames to be encoded" were set to 1. The desiredcompression level was achieved by tuning the quantizationparameter (QP). All other parameters were set to their defaultvalues.The WAC leverages the orthonormal wave atom transform.This non-adaptive multi-resolution transform has good space-frequency localization and its orthonormal basis is suitablefor sparsifying holographic signals. Basically this codec isbased on a JPEG 2000 coding architecture where the CDF 9/7wavelet transform is replaced by the 2D wave atom transformwhere the spatial footprint of each atom scales paraboli-cally across resolutions, while the quantization and EmbeddedBlock Coding by Optimizated Truncation (EBCOT) [37] arefurther deployed. EBCOT code blocks of size 128 ×
128 pixelsare issued.All three encoders compressed the 8-bit quantized real andimaginary parts of each hologram separately. The hologramswere compressed at bitrates 0.25 bpp, 0.5 bpp, 0.75 bpp and 1.5 bpp. These bitrates were determined via a series ofmock-up tests where the holograms were compressed at 9different bit-depths between 0.15 bpp to 4 bpp and their visualappearance were tested on 2D and light field setups. Also onthe holographic display, 3 sample holograms were tested forall the 9 bit-depths and the others were verified for the chosenbit depths. The goal was to ensure that the distortion levelsresulted in broad range of visual quality levels ranging fromvery poor to imperceptible.The full set of the test holograms along with their acquiredquality scores and other related data to these experiment arepublicly available at: http://data.etrovub.be/holodb.IV. T
EST METHODOLOGY
A. Generic procedure for subjective quality assessment
Holographic modalities and in extension, plenoptic modal-ities, pose specific challenges as it concerns evaluating theirvisual quality. This is due to the fact that the plenoptic function"allows for the reconstruction of every possible view, at everymoment, from every position, at every wavelength within thespace-time wavelength region under consideration" [38]. Asa consequence, to meet its time and resource constraints, asubjective experiment has to be limited to evaluate only a nec-essary subset of this 7D space. For instance, several subjective
UBMITTED TO THE JOURNAL OF IEEE TRANSACTIONS ON MULTIMEDIA,20 SEPTEMBER 2019 6
Table II: Details of the test conditions and the gathered scores per display setup.(* The hologram "Mermaid" was reconstructedat 1 focal distance and "Chess" at 3. Although the average number of reconstructions per hologram is equal to 2.)
Setup No. Tested Objects Distortions Dist. Levels Perspective Recon. Distance Total Conditions Scores per ConditionOptical
Light field
Regular 2D
Figure 6: Subjective testing experimental setups for 2D display (a), light-field display(b) and the holographic display setup (c)quality assessment approaches have been reported for 4D lightfields. At IEEE ICME 2016 a Grand Challenge on Light FieldCompression was organized [39] deploying a double stimuluscontinuous quality scale (DSCQS) methodology [40]. In thissolution uncompressed and decoded views are shown side-by-side on a high-end monitor while selecting a limited set ofviews and focus points per light field. In the context of theJPEG Pleno Light Field Coding standardization effort and a as-sociated Grand Challenge on Light Field Coding organized atIEEE ICIP 2017, a double stimulus comparison scale (DSCS)methodology was employed with side-by-side rendering of thelight field as a pseudo video sequence and using a discretequality scale ranging from -3 to 3 [41], [42]. Viola et al. [43]assessed the impact of an interactive approach to determineperceived quality as such enabling to evaluate a larger fractionof the light field compared to the passive approach of [42]. Thesame authors applied the advocated solution also to evaluate alarger set of light field compression techniques [44]. For pointclouds data few subjective quality assessment experimentshave been also conducted. To evaluate point cloud compressiontechniques, Javaheri et al. [45] rendered a video sequenceby having a virtual camera spiralling around the point cloudobject. The Double Stimulus Impairment Scale (DSIS) [40]methodology was adopted and the video sequences of theimpaired and original point cloud were shown sequentially. Asimilar procedure was applied for the evaluation of point clouddenoising algorithms [46]. For holographic data few earlierefforts took place, several open access test data bases havebeen proposed; such as: the B-Com Repository [47], [48], ERCInterfere I [49], II [27] and III [14], and EmergImg-HoloGrailv1 and v2 [10]. Recently, Amirpourazarian et al. presenteda methodology to evaluate perceptual quality of compressedholograms on a 2D display [50].As mentioned in section I, the testing method should beadapted to the specific limitations and different technicalrequirements of each display type. The holograms were shownmainly following the procedure for DSIS. In our method the reference and distorted stimuli are sequentially shown tothe subject and then the subject scores the second stimulus(impaired version) based on the first (reference). The hologramsequences were shown in a fully randomized order. Thepresentation order was also randomized for each subject.The scoring procedure was followed by the standard oneproviding 5 quality scales. Depending on the perceived mis-match, subject chooses a quality number from 1 to 5 repre-senting one of the impairment scales: Very Annoying, An-noying, Slightly Annoying, Perceptible but not Annoying, andImperceptible. The testlab conditions corresponded to ITU-R BT.500-13 recommendations [40] and recommendationsdescribed in Annex B of ISO/IEC 29170-2 (AIC Part-2).
B. Subjective quality assessment on holographic display
The subjective test on the holographic display was con-ducted in the photonics laboratory of the Institute of Mi-cromechanics and Photonics of Warsaw University of Technol-ogy. The holograms were shown mainly following the DSISprocedure. From each synthetic aperture hologram a sub-holograms of 2048 × UBMITTED TO THE JOURNAL OF IEEE TRANSACTIONS ON MULTIMEDIA,20 SEPTEMBER 2019 7 performed in two days such that each subject participated inonly 2 sessions per day. A compulsory minimum of 5 minuterest was facilitated by the test operator before starting the nextsession. The maximum rest time was not limited and subjectshad the freedom to take larger recuperation periods in casethey felt it to be necessary.
C. Subjective quality assessment on light field display
For the Light-field display, again the DSIS method wasimplemented following the ITU-BT.500-13 recommenda-tions [40]. Although, in this case, the display provides a wideangle, simultaneous rendering of multiple views and eachsubject was required to watch and score the center and right-corner view of each displayed hologram-pair. To facilitate arepeatable procedure, the places where subjects have to standto see the required views, were marked on the ground. Thedistance from the screen was chosen . times the height of thescreen. For each tested hologram, the subject starts standingin the center-position and the operator displayed the referenceand impaired holograms sequentially. After recording thescore, the subject moved to the right-corner position and againboth reference and impaired holograms were displayed by theoperator followed by the scoring. According to Table II thenumber of test-conditions per subject was twice the numberof test-conditions per subject in the holographic setup. Thisis due to the fact that for each hologram, test subjects scoredthe visual quality at two different reconstructions distancesfor the light field display. The test in this setup was conductedin 2 sessions with a target duration of 20 minutes. Since thesubjects were required to stand and move multiple times todesignated positions during the test, at least 1 hour rest wasconsidered before starting the second session. D. Subjective quality assessment on 2D display
For the 2D setup, see Fig. 6.a, each reconstructed hologramwas shown for the 2 perspective positions corresponding tothe one for the light field display and holographic displayand two reconstruction distances corresponding to the lightfield display test. The reconstructed reference and impairedholograms were displayed side by side reducing the test timeper subject by half.
E. Training of test subjects
For each setup, 40 subjects participated. From the total of120 participants, the number of female and male participantswere 54 and 66 respectively. Their age was between 18 to30 years old. Prior to the test, subjects were required to passthe Snellen visual acuity test. Though, all the content shownto the subjects was monochromatic, the Ishihara test to detectthe colorblindness was performed as well. Prior to the firsttest session in each setup, a 5 minute training session wasconducted where the test and scoring procedure was explainedand rehearsed. (a) Light-Field display(b) 2D display(c) Holographic display
Figure 7: Histogram of Z-scores per display setup - calculatedper condition from the raw scores before outlier removal andaveraging. The histograms represent the underlying distribu-tions of raw scores and are thus directly comparable. Thepercentage of Z-scores, which falls within 1 and 2 standarddeviation(s) from their mean (their MOS after outlier removal)is shown on the graphs.V. R
ESULTS AND A NALYSIS
In this section, we provide the results of our subjectiveexperiments and further investigate various aspects of theoutcomes, potential similarities and correlations among thegathered scores from the three testing setups.
A. Reliability analysis of the obtained MOS
First, a reliability analysis is performed on the acquiredopinion scores for the three setups. Before performing any
UBMITTED TO THE JOURNAL OF IEEE TRANSACTIONS ON MULTIMEDIA,20 SEPTEMBER 2019 8 post-processing on the scores and calculating the Mean Opin-ion Score (MOS) for each test condition (see proceduredescribed in Sec. V-B), it is important to check whetherthe average is a reliable representative of the underlyingdistribution per condition. To determine the MOS reliability,one should ideally identify the distribution model of the data.Though, considering our limited sample size (20 scores percondition, from each setup) conventional statistical modelingmay not necessarily reach to a conclusive result. Instead, akurtosis analysis has been recommended in standards likeITU.BT500.13 [40], where a score distribution with a kurtosisvalue of 2 to 4 is interpreted as a representative of thenormal probability model. However, this is a vague and flawedassumption. It is indeed correct that the kurtosis of a normaldistribution model is equal to 3, though mathematically thisis a necessary but not sufficient condition. Moreover, bydefinition its only unambiguous interpretation is in terms ofdistribution tail extremity [51]. Nonetheless, no score set (percondition) in our dataset showed any irregular kurtosis value.Next, we seek to answer two questions: (1) Whether thesubject scores for a specific test condition reach a consensusabout the visual quality score for this test condition or not?(2) If the answer to the first question is positive, to whatextend can that consensus be represented by the mean ofthese scores? To compactly address both, first we standardizethe scores per condition. Z-scores are calculated where eachscore is normalized by the mean and standard deviation ofthe scores for the same test condition. The advantage of Z-scores is that their normalization enables direct comparisonof individual scores across all conditions and even differentsetups. Nonetheless, the Z-score value does not provide anyinformation about the actual visual quality level. It gives thedistance of each individual score from the average opinionscore (in units of standard deviation). This way a histogramof all scores for a particular setup (Fig. 7) gives an abstractview on the overall agreement of test subjects. Notably, forall setups, a significantly good agreement is available aroundthe mean opinion values, such that more than . and . of the individual scores fall within only 1 and 2standard deviations(s) away from their corresponding mean,respectively. (The corresponding values for a perfect normaldistribution are . and . ). Based on this and the factthat no specific skewness can be seen around the tails of showndistributions, we believe our MOS values can appropriatelyrepresent opinions of the majority of tested subjects. B. Statistical processing of results
The distributions of Fig. 7 shows that a very small portionof scores per setup are more than 4 standard deviations awayfrom the average scores per condition. Therefore, an outlierdetection and removal was performed on the test results.Following the procedure used in [52] and [12], the th ( Q ) and th ( Q ) percentiles were calculated. A score u was considered as an outlier if u > Q w ( Q − Q or u < Q − w ( Q − Q , where w was the maximum whiskerlength. w = 1 . for normally distributed data corresponds to . coverage, which was utilized in the experiment. Our (a) MOS
LFfront vs MOS
LFback (b)
MOS Dfront vs MOS Dback (c)
MOS
LFfront vs MOS
LFback (d)
MOS Dfront vs MOS Dback
Figure 8: Overall comparison of the Front-Focus MOS versusBack-Focus MOS for the light field display, (a) and (c), andthe 2D display, (b) and (d). The results shown in (a) and (b) foreach depth are averaged over center and corner perspectives.The raw data is shown color coded for both cases in (c) and (d).The indicated lines in (a) and (b) are 4th-degree polynomial fitlines for the data with indices of the sorted front focus MOSand sorted back focus MOS, respectively.results also showed that no test subject had more than outlier scores. Consequently, no test subjects were removedfrom the dataset. After removing the outlier scores, the averageof the remaining scores for a particular test condition wascombined into the final MOS.
C. MOS analysis based on reconstruction focal-point
First, the MOS values at different reconstruction distanceswere evaluated for light field and 2D displays. Fig. 8 showsthe overall comparison between front and back focus MOS,while each MOS is averaged between the two perspectives(Center and Right-Corner views). It is obvious that the MOSfrom both depths are very closely following the same trend.Nonetheless, the non-averaged MOS are also visualized inthe scatter plots of Fig. 8(c, d). Therein points are coloreddifferently by perspective. At this point, results do not showany meaningful differences. Therefore, to limit the degrees offreedom for our analysis, we use in the next subsections theMOSes, which have been averaged over the focal distances.This means the number of MOSs for light field and 2D setupswill be equal to the ones from the holographic setup (96 scoresper perspective and a total of 192 scores per setup).
UBMITTED TO THE JOURNAL OF IEEE TRANSACTIONS ON MULTIMEDIA,20 SEPTEMBER 2019 9 (a)
MOS
OPTc vs MOS
OPTr (b)
MOS
LFc vs MOS
LFr (c)
MOS Dc vs MOS Dr (d) MOS
OPTc vs MOS
OPTr (e)
MOS
LFc vs MOS
LFr (f)
MOS Dc vs MOS Dr Figure 9: Overall comparison of the center view MOS versus right corner view MOS for (9a, d) Holographic display, (b,e)Light-field display and (c,f) 2D regular display. In (a,b,c) additionally 4th-degree polynomial fit lines are shown for the datawith indices of the sorted center view MOS.
D. MOS analysis based on perspective
Next, the correlation between the scores for the two testedperspectives are evaluated. Fig. 9 shows per setup the compar-ison of the MOS from the center view with the right-cornerview. For each setup, first the center scores were sorted andthe sorting indices were used to plot the corner view MOS.The confidence intervals for each perspective are shownas well. To avoid clutter and to further clarify the trend, only4th degree polynomial fit lines for the mentioned data areshown in Fig. 9 (a,b,c). To provide more detail, scatter plotsof the center vs right corner MOSes are shown in Fig. 9 (d,e,f).The score plots clearly show a distinct trend across the setupswhere central views regularly obtain a higher MOS comparedto the corner views of the same hologram. However, the scoredifference evolves across the quality range. More specifically,for all setups, the corner view MOS for high quality holograms(holograms with center view MOS higher than 3.5) remainswithin the confidence interval fits of the center view MOS.On the other hand for holograms in the lower end of qualityrange, the differences increase. This is perfectly in line withthe expected behaviour of how some encoders compress theholograms. When performing lossy compression the generalobjective is to weight the transform components in the space-frequency domain higher, which carry more visually importantinformation. However, if very high compression ratios are demanded, this will translate into complete elimination of theweakest coefficients. This leads, in the case of the chosenWAC variant, to an introduction of overlapping first diffractionorders by imperfect coefficient cancellation, which is morepronounced away from the center. In the case of the otherselected methods, it leads to an elimination of high frequencycomponents, which correspond to high diffraction angles (cor-ner view information). The MOS variations experimentallyreveals this shortcoming of the current holographic encoders.The scatter plots show furthermore that there are some casesthat do not follow this difference trend. In some extreme casesthe center MOS is 1.5 points higher than the corner MOS.
E. Inter-setup comparison of results
In this section, the MOS results obtained with the differentdisplay systems are compared. First, the overall trend of thescores is evaluated. Thereafter, a more detailed analysis isperformed related to the influence of the characteristics of theencoded objects and the bit-depths used to encode them in thisexperiment.Fig. 10 shows the inter-setup comparison for the center-view. Similar to Fig. 9, the MOS results for the optical setup(a,b) and for the light field display (c) were sorted and theirorder was used to plot the other corresponding MOS for theother setup. The lines (solid and dashed) depict again the 4th-degree polynomial fits for the corresponding data and their
UBMITTED TO THE JOURNAL OF IEEE TRANSACTIONS ON MULTIMEDIA,20 SEPTEMBER 2019 10 (a)
MOS
OPTc vs MOS
LFc (b)
MOS
OPTc vs MOS Dc (c) MOS
LFc vs MOS Dc (d) MOS
OPTc vs MOS
LFc (e)
MOS
OPTc vs MOS Dc (f) MOS
LFc vs MOS Dc Figure 10: Center-view MOS comparison for holographic versus Light Field display (a,d), holographic versus 2D display (b,e)and Light Field versus 2D display (c,f).Table III: Calculated coefficients of the fit functions, which enable to map the gathered MOS from different setups into anotherfor the center and the corner views. Pearson and Spearman correlation coefficients are provided to evaluate the accuracy ofthe functions. As for the robustness, the last column shows the maximum absolute error in unit score for the predicted fit, ifone of the test subjects changes a score for a condition with ± unit score. p ( x ) = p x + p x + p x + p x + p . Before Fit After Fit Error p p p p p Pearson Spearman Pearson SpearmanCenter View LF → OPT → OPT → LF Corner View LF → OPT → OPT → LF confidence intervals. The data shown in Fig. 10(a, d), representthe comparison between the optical and light field setup.Interestingly, a specific gap exists between the two setups.For each hologram the MOS obtained for the optical setupis typically higher than the MOS for the light field setup -except in the very low quality range. A similar trend can beobserved in Fig. 10 (b,e) where the optical setup was comparedwith a 2D display setup. However, the gap for the mid-rangequality holograms (scores . − ) is slightly larger now. Onthe other hand, the graphs shown in Fig.10(c, f), demonstratea rather close agreement between the MOS of the light fieldand 2D display setups. A comparison for the right cornerview perspective is provided in Fig. 11 and the right cornerview scores follow closely the trend found for the center view.However, the level of disagreement between some individual light field and 2D display scores is slightly increased for thecorner-views. Unfortunately, we do not have an immediateexplanation for this phenomena.Additionally, for the cases where quality scores of a realholographic setup are not available, while having access tothe 2D or light field scores; one can use the fit-functionsevaluated and shown as blue lines in the scatter plots of Fig. 10(d,e) and Fig. 11 (d,e) to predict the scores for the same datareconstructed in optical setup.Table III shows the coefficients for the 4th-degree poly-nomials (Fig. 10 and Fig. 11), which are best fit in a least-squares sense, for the three comparisons between each pairof the display setups. During our experiments, the 4-th degreepolynomial showed the lowest regression error while not over-fitting the data when comparing the fitting behaviour of poly- UBMITTED TO THE JOURNAL OF IEEE TRANSACTIONS ON MULTIMEDIA,20 SEPTEMBER 2019 11 (a)
MOS
OPTr vs MOS
LFr (b)
MOS
OPTr vs MOS Dr (c) MOS
LFr vs MOS Dr (d) MOS
OPTr vs MOS
LFr (e)
MOS
OPTr vs MOS Dr (f) MOS
LFr vs MOS Dr Figure 11: Right corner-view MOS comparison for holographic versus Light Field display (a,d), holographic versus 2D display(b,e) and Light Field versus 2D display (c,f).nomials of degree 1 to 7. In the same table, the Pearson andSpearman correlation coefficients are shown before and afterapplying each fit function to the data. A logical concern, whicharises here, is the robustness of the provided fit functions.For example, in the case a test subject changes its opinionabout a hologram. The last column of the Table III representsthe maximum possible change of the fitted values, if a singlesubject changes the score for a test condition by ± unit. Thereported errors are in unit scores as well.These results indicate that a high correlation exists betweenthe MOS obtained for the three display systems and, inparticular, after polynomial fitting. Both Pearson and Spear-man correlation coefficients are very large in the latter case,thus underlining the predictive power of 2D and light fielddisplays with respect to a holographic setup. Nonetheless, itis important to understand these fits cannot be transferred toother display setups and a calibration process will always beneeded. When looking at the non-fitted MOS, it is interestingto observe that the 2D and LF displays are more sensitive toartifacts than the holographic display. This is partially relatedto the higher quality of the issued 2D and light field displays,but also due to the fact they are displaying numericallyreconstructed holograms which contain more coherent specklenoise than the content rendered on the holographic display,for which a partially coherent LED illumination reduced thiseffect. Also non-optimal optics further reduce naturally the amount of coherent speckle noise. The test subjects werenot familiar with the phenomenon of speckle noise and wereinstructed to ignore it for the 2D and LF displays. Though, itmight have influenced their scoring.Apart from the overall inter-setup comparisons, it is in-teresting to analyse the influence of the tested hologram onthe scoring behaviour of the test subjects: do responses tothe test display systems differ for different holograms? Forreasons of brevity, only the difference range (dashed line), the25%-75% interquartile interval and the median difference isprovided. Fig. 12 represents the inter-setup MOS differencesper test hologram (for all test conditions), seperately forthe center (a,b,c) and right-corner (d,e,f) views. Here, weconsider the median difference between the MOS of eachpair of setups as an indicator of the difference between thescores (shown as red line inside each box). The smaller theinterquartile range for each boxplot, the higher the certainty onthe difference-level (red-line) of the opinions for that object.For example, in Fig. 12.a, the MOS of the optical setup forall distorted versions of "Mermaid" is ≈ . larger than theMOS for the light field setup. Considering the small size of theinterquartile-range ≈ . , one may conclude that the showndifference almost equally persists across all the distortiontypes and distortion levels. When comparing Fig.12.(a,b,c),the general trend related to the MOS differences for thedifferent setups (shown in Fig. 10), persists for each tested UBMITTED TO THE JOURNAL OF IEEE TRANSACTIONS ON MULTIMEDIA,20 SEPTEMBER 2019 12 (a) ( MOS
OPTc − MOS
LFc ) (b) ( MOS
OPTc − MOS Dc ) (c) ( MOS
LFc − MOS Dc ) (d) ( MOS
OPTr − MOS
LFr ) (e) ( MOS
OPTr − MOS Dr ) (f) ( MOS
LFr − MOS Dr ) Figure 12: Box plots of the MOS difference per hologram. First-row boxplots(a,b,c) corresponds to the MOS differences forcenter view and second row (d,e,f) show the MOS differences collected from right-corner view. Column-wise, the boxplotsrepresent MOS diffrences between holographic (optical setup) - light field display, holographic - 2D regular display, and lightfield-2D display, respectively.hologram individually. Thereby the MOS values between non-holographic displays are spread less compared to the MOSof the holographic display. The MOS results for "Chess" arethe only results that show a rather stable behaviour across allsetups.In Fig. 13, another categorization of the MOS differencesbetween the three setups and two perspectives is shown.Here, the MOS obtained from the holograms with the samecompression level (quality range) are compared across setups.When considering the median differences (red lines), a rathersimilar trend is recognizable in Fig. 13 (a,b,d, and e), wherethe score gap between the holographic display and 2D or LFdisplays for the bit-depths 0.5 and 0.75 bpp are larger; whilethe level of disagreement is smaller for the lowest and thehighest bit-depth. The certainty level of these results (referringto the size of the interquartile range) increases for higher bit-depths as well. In the case of a direct comparison of LF and2D display the differences are statistically not relevant.VI. C
ONCLUSIONS
In this paper, we reported the results of a series of com-prehensive subjective experiments where, for the first time, aset of digital holograms was designed and created to evaluatevarious aspects of macroscopic holography. For each hologram 12 distorted versions were generated by compression at differ-ent bit-rates using state-of-the-art holographic encoders. Threeseparate subjective experiments were designed and imple-mented utilizing a holographic display, a light-field, and a 2Dstandard monitor. For the subjective tests, a double stimulus,multi-perspective, multi-depth subjective testing methodologywas designed and adapted to the characteristics of the utilizeddisplays. A total of 120 human subjects participated in theexperiments. The acquired quality scores of the reconstructedholograms were compared based on perspective and focaldistance. Our results showed no explicit distinction betweenthe scores of holograms when different parts of the encodedobjects were in focus. However, with a change of perspectivethere was a consistent gap between the rated visual qualities.The corner-view generally scored lower than the center-view,especially for mid-range and high distortion levels. Further,we compared the scores obtained from a holographic displaywith the scores from the Light-field and 2D displays andidentified another rather consistent and distinctive gap. Ourresults show that the same distorted holograms rendered onholographic displays appear less distorted to the human eyethan it is the case for light-field or 2D display. However, itwas demonstrated that the scores on different displays arehighly correlated and follow a consistent trend through thequality range. This indicates that numerically reconstructed
UBMITTED TO THE JOURNAL OF IEEE TRANSACTIONS ON MULTIMEDIA,20 SEPTEMBER 2019 13 (a) ( MOS
OPTc − MOS
LFc ) (b) ( MOS
OPTc − MOS Dc ) (c) ( MOS
LFc − MOS Dc ) (d) ( MOS
OPTr − MOS
LFr ) (e) ( MOS
OPTr − MOS Dr ) (f) ( MOS
LFr − MOS Dr ) Figure 13: Box plots of the MOS difference per compression level. The first and second rows shows the MOS differences forthe center and right-corner view, respectively.holograms displayed on light field or 2D displays allow forappropriate predictions on the perceptual visual quality ofholographic displays. For completeness, we also provided fit-functions which map scores from different setups into oneanother. Finally, we are hoping that the provided resultsplus the annotated database of our holograms, which arepublicly available, facilitate a reliable test-bed for designingor improving available holographic processing methods andplenoptic quality metrics, as well as systematic benchmarkingoperations for digital holograms.A
CKNOWLEDGMENT
Research for this paper received funding from the EuropeanResearch Council under the European Union’s Seventh Frame-work Programme (FP7/2007-2013)/ERC Grant Agreementn.617779 (INTERFERE), the Cross-Ministry Giga KOREAProject (GK19D0100, GigaKOREA) and Warsaw Universityof Technology.The models “Perforated Ball“ and “Biplane” are courtesyof Manuel Piñeiro from GrabCad.com and ScanLAB Projects,respectively. R
EFERENCES[1] B. Lee, “Three-dimensional displays, past and present,”
Physics Today ,vol. 66, no. 4, p. 36, 2013. [2] D. Blinder, A. Ahar, S. Bettens, T. Birnbaum, A. Symeonidou,H. Ottevaere, C. Schretter, and P. Schelkens, “Signal processingchallenges for digital holographic video display systems,”
SignalProcessing: Image Communication
Opt. Express
Journal of Information Display ,vol. 18, no. 1, pp. 1–12, Jan. 2017. [Online]. Available: https://doi.org/10.1080/15980316.2016.1255672[5] Y. Pan, J. Liu, X. Li, and Y. Wang, “A Review of Dynamic HolographicThree-Dimensional Display: Algorithms, Devices, and Systems,”
IEEETransactions on Industrial Informatics , vol. 12, no. 4, pp. 1599–1610,Aug. 2016.[6] T. Sugie, T. Akamatsu, T. Nishitsuji, R. Hirayama, N. Masuda,H. Nakayama, Y. Ichihashi, A. Shiraki, M. Oikawa, N. Takada,Y. Endo, T. Kakue, T. Shimobaba, and T. Ito, “High-performanceparallel computing for next-generation holographic imaging,”
NatureElectronics
Computer Holography:Acceleration Algorithms and Hardware Implementations
Applied Optics , vol. 57, no. 17, pp. 4930–4942, Jun. 2018,wOS:000434872300027.[9] J. P. Peixeiro, C. Brites, J. Ascenso, and F. Pereira, “Holographic DataCoding: Benchmarking and Extending HEVC With Adapted Trans-
UBMITTED TO THE JOURNAL OF IEEE TRANSACTIONS ON MULTIMEDIA,20 SEPTEMBER 2019 14 forms,”
IEEE Transactions on Multimedia , vol. 20, no. 2, pp. 282–297,Feb. 2018.[10] M. V. Bernardo, P. Fernandes, A. Arrifano, M. Antonini, E. Fonseca,P. T. Fiadeiro, A. M. G. Pinheiro, and M. Pereira, “Holographic repre-sentation: Hologram plane vs. object plane,”
Signal Processing: ImageCommunication
ETRI Journal ,pp. 93–108, 2019.[12] A. Ahar, D. Blinder, T. Bruylants, C. Schretter, A. Munteanu, andP. Schelkens, “Subjective quality assessment of numerically recon-structed compressed holograms,” in
Applications of Digital Image Pro-cessing XXXVIII , vol. 9599. United States: SPIE, 2015, pp. 95 990K–95 990K–15.[13] A. Symeonidou, D. Blinder, B. Ceulemans, A. Munteanu, andP. Schelkens, “Three-dimensional rendering of computer-generated holo-grams acquired from point-clouds on light field displays,” in
SPIE Optics+ Photonics, Applications of Digital Image Processing Xxxix , A. G.Tescher, Ed., vol. 9971. Bellingham: Spie-Int Soc Optical Engineering,2016, p. 99710S, wOS:000390023100024.[14] A. Symeonidou, D. Blinder, and P. Schelkens, “Colour computer-generated holography for point clouds utilizing the phong illuminationmodel,”
Opt. Express
Opt. Express
Proc. SPIE , vol. 10834, 2018, pp. 10 834 – 10 834 – 6.[Online]. Available: https://doi.org/10.1117/12.2320114[17] X. Li, J. Liu, J. Jia, Y. Pan, and Y. Wang, “3d dynamic holographic dis-play by modulating complex amplitude experimentally,”
Optics express ,vol. 21, no. 18, pp. 20 577–20 587, 2013.[18] T. Kozacki, M. Chlipala, and P. L. Makowski, “Color fourierorthoscopic holography with laser capture and an led display,”
Opt. Express
Journal of the Optical Society of America A ,vol. 13, no. 3, pp. 470–473, Mar. 1996. [Online]. Available:http://josaa.osa.org/abstract.cfm?URI=josaa-13-3-470[20] J. W. Goodman,
Introduction to Fourier Optics
Proc. SPIE , vol. 9970, 2016, Conference Proceed-ings.[24] D. Claus, D. Iliescu, and P. Bryanston-Cross, “Quantitative space-bandwidth product analysis in digital holography,”
Applied optics ,vol. 50, no. 34, pp. H116–H127, 2011.[25] G. Stroke, “Lensless fourier-transform method for optical holography,”
Appl. Phys. Lett. , vol. 10, no. 6, pp. 201–203, 1965.[26] M. Guizar-Sicairos, S. T. Thurman, and J. R. Fienup, “Efficient subpixelimage registration algorithms,”
Opt. Lett.
Proc. SPIE , vol.9896, 2016, p. 98960F. [Online]. Available: http://dx.doi.org/10.1117/12.2225201[28] J. W. Goodman,
Introduction to Fourier optics . Roberts and CompanyPublishers, 2005.[29] K. Matsushima, “Shifted angular spectrum method for off-axisnumerical propagation,”
Opt. Express
Proceedings of the IEEE , vol. 90, no. 8, pp. 1336–1357,Aug. 2002.[31] P. Schelkens, A. Skodras, and T. Ebrahimi,
The JPEG 2000 suite . JohnWiley & Sons, 2009, vol. 15.[32] G. J. Sullivan, J.-R. Ohm, W.-J. Han, and T. Wiegand, “Overviewof the High Efficiency Video Coding (HEVC) Standard,”
IEEETransactions on Circuits and Systems for Video Technology , vol. 22,no. 12, pp. 1649–1668, Dec. 2012. [Online]. Available: http://ieeexplore.ieee.org/document/6316136/[33] T. Birnbaum, A. Ahar, D. Blinder, C. Schretter, T. Kozacki, andP. Schelkens, “Wave Atoms for Lossy Compression of Digital Holo-grams,” in
Signal Processing-Image Communication ,pp. 112–133, Feb. 2015, wOS:000350078800010.[36] Fraunhofer, “HEVC reference repository,” https://hevc.hhi.fraunhofer.de/svn/svn{_}HEVCSoftware/branches/HM-16.18-dev, 2018 (accessedJuly 1, 2018).[37] D. Taubman, “High performance scalable image compression withEBCOT,”
IEEE Transactions on Image Processing , vol. 9, no. 7, pp.1158–1170, Jul. 2000.[38] E. H. Adelson and J. R. Bergen, “The Plenoptic Function and the Ele-ments of Early Vision,” in
Computational Models of Visual Processing .MIT Press, 1991, pp. 3–20.[39] I. Viola, M. ˇReˇrábek, T. Bruylants, P. Schelkens, F. Pereira, andT. Ebrahimi, “Objective and subjective evaluation of light field imagecompression algorithms,” in , Jul. 2018, pp. 1–6.[42] ISO/IEC JTC1/SC29/WG1, “JPEG Pleno Call for Proposals on LightField Coding,” Jan. 2017, wG1N74014, 74th JPEG Meeting, Geneva,Switzerland.[43] I. Viola, M. ˇReˇrábek, and T. Ebrahimi, “Impact of interactivity on theassessment of quality of experience for light field content,” in , May 2017, pp. 1–6.[44] ——, “Comparison and Evaluation of Light Field Image Coding Ap-proaches,”
IEEE Journal of Selected Topics in Signal Processing , vol. 11,no. 7, pp. 1092–1106, Oct. 2017.[45] A. Javaheri, C. Brites, F. Pereira, and J. Ascenso, “Subjective andobjective quality evaluation of compressed point clouds,” in ,Oct. 2017, pp. 1–6.[46] ——, “Subjective and objective quality evaluation of 3d point clouddenoising algorithms,” in , Jul. 2017, pp. 1–6.[47] A. Gilles, P. Gioia, R. Cozot, and L. Morin, “Hybrid approach forfast occlusion processing in computer-generated hologram calculation,”
Applied optics , vol. 55, no. 20, pp. 5459–5470, 2016.[48] ——, “Computer generated hologram from multiview-plus-depth dataconsidering specular reflections,” in . IEEE, 2016, pp. 1–6.[49] D. Blinder, A. Ahar, A. Symeonidou, Y. Xing, T. Bruylants, C. Schret-ter, B. Pesquet-Popescu, F. Dufaux, A. Munteanu, and P. Schelkens,“Open Access Database for Experimental Validations of HolographicCompression Engines,” in . New York: Ieee, 2015,wOS:000375091800066.[50] H. Amirpourazarian, E. Fonseca, M. Pereira, and A. Pinheiro, “QualityAssessment of Holographic Data,” Doc. ISO/IEC JTC 1/SC 29/WG1M84117, 84th Meeting, Brussels, Belgium, Geneva, Jul. 2019.[51] P. H. Westfall, “Kurtosis as peakedness, 1905–2014. rip,”
The AmericanStatistician , vol. 68, no. 3, pp. 191–195, 2014.[52] P. Hanhart, M. Rerabek, F. De Simone, and T. Ebrahimi, “Subjectivequality evaluation of the upcoming hevc video compression standard,”in
SPIE Optical Engineering+ Applications . International Society forOptics and Photonics, 2012, pp. 84 990V–84 990V.
UBMITTED TO THE JOURNAL OF IEEE TRANSACTIONS ON MULTIMEDIA,20 SEPTEMBER 2019 15
Ayyoub Ahar received his B.Sc. degree in appliedphysics from PNU and the M.Sc. degree in IT-multimedia computing from Multimedia University,Malaysia. He is currently pursuing the Ph.D. de-gree in electrical engineering with Vrije UniversiteitBrussel (VUB), Belgium. He is a member of the IN-TERFERE, a multidisciplinary research group at theDepartment of Electronics and Informatics, VUB.He holds a full scholarship as a part of EU-ERCConsolidator Grant focusing on digital holography.He is also a Researcher at IMEC, Leuven, Belgium.His current research interests include digital image and video processing,complex data analysis, bio-inspired computing, perceptual quality assessmentand standardization with an emphasis on emerging 3D data modalities, inparticular digital holography and light field imaging.
Maksymilian Chlipala received his PhD degree inarea of construction and exploitation of machinesfrom the Faculty of Mechatronics, Warsaw Uni-versity of Technology, Warsaw, Poland, in 2019.His main research interests are digital holography,holographic displays, Spatial Light Modulators andspeckle reduction with LED sources.
Tobias Birnbaum received his B.Sc. degree inApplied Natural Science (2012) and his Dipl. in Ap-plied Mathemathics (2016) from TU BergakademieFreiberg, Germany. He is currently working towardsa PhD in the field of signal processing of digitalholograms at Vrije Universiteit Brussels, Belgiumas part of the INTERFERE EU465 project. Hismain research interests are compression and post-processing, such as denoising, of digital holograms,as well as compressed sensing, and dictionary learn-ing.
Weronika Zaperty received the M.Sc. degree fromWarsaw University of Technology, Warsaw, Polandin 2011, where she is currently working toward thePh.D. degree in the field of color digital hologra-phy. Her scientific research are related to digitalholography, holographic displays and holographicinterferometry.
Athanasia Symeonidou received a B.Sc. degree inPhysics in 2009 and a M.Sc. degree in ElectronicPhysics in 2012 from Aristotle University of Thes-saloniki, Greece. She currently works towards herPh.D. in Engineering Sciences at Vrije UniversiteitBrussel, Belgium, as a researcher on the ERC projectINTERFERE. Her research interests are 3D imaging,digital holography and signal processing, focusingon efficient algorithms for generation and display ofhigh quality computer-generated holograms.
Tomasz Kozacki received his Ph.D. in the fieldof Photonics at Warsaw University of Technologyin 2005 and habilitation in 2013, where he alsoworks as a professor. His scientific research arerelated to digital holography, holographic displays,holographic microscopy, computational diffractionand optical diffraction tomography. He has authoredand co-authored more than 30 scientific journal pub-lications and more than 50 conference proceedings.
Malgorzata Kujawinska
PhD DSc., SPIE Fellow,Full Professor of applied optics at Warsaw Uni-versity of Technology. Expert in full-field opticalmetrology, development of novel photonics measure-ment systems and holographic displays, 3D quan-titative imaging in multimedia and biomedical en-gineering. Author of one monograph, several bookchapters and more than 250 papers in internationalscientific journals. She had been the SPIE Presidentand vice-President of European Technology PlatformPhotonics21. The recipient of SPIE 2013 Chandra S.Vikram Award in Optical Metrology.