Inter national J our nal of Electrical and Computer Engineering (IJECE) V ol. 9, No. 4, August 2019, pp. 2394 2402 ISSN: 2088-8708, DOI: 10.11591/ijece.v9i4.pp2394-2402 r 2394 Computer vision based 3D r econstruction : A r e view Hanry Ham, J ulian W esley, Hendra Computer Science Department, School of Computer Science, Bina Nusantara Uni v ersity , Indonesia Article Inf o Article history: Recei v ed Jan 15, 2018 Re vised Jan 23, 2019 Accepted Mar 4, 2019 K eyw ords: 3D alignment 3D point clouds 3D reconstruction ABSTRA CT 3D reconstruction are used in man y fields starts from the object reconstruction such as site, cultural artif acts in both ground and under the sea le v els, medical imaging data, nuclear substantional. The scientist are beneficial for these task in order to learn, k eep and better visual enhancement into 3D data. In this paper we dif ferentiate the algorithm used depends on the input image : single still image , RGB-Depth image, multiperspecti v e of 2D images, and video sequences. The prior w orks also e xplained ho w the 3D reconstruction perform in man y fields and using v arious algorithms. Copyright c 2019 Institute of Advanced Engineering and Science . All rights r eserved. Corresponding A uthor: Hanry Ham, Computer Science Department, School of Computer Science, Bina Nusantara Uni v ersity , Jakarta, 11480 - Indonesia. Email: hanry .ham@binus.edu 1. INTR ODUCTION 3D Reconstruction task is one of the interesting task that meet i ts maturity already . These can be seen from the commercial products such as product from Agisoft and Pix4D that are capable of produced high quality of lar ge scal e 3D models. Furthermore, the hardw are such as the computer vision has been de v eloped and impro v e si nce then. There are some setup camera introduced in the research such as stereo camera and Kinect. In addition to the vision setup, ki nect camera sho ws a great positi v e feedback from the r esearchers, pro v ed by common vision setup that can be found in the literature re vie w . Not only that, stereo camera setup can be found among the literature re vie w . In addition to the stereo camera, custom stereo camera are quite popular among the researchers by combining tw o equals web camera that positioned by period of distance. The algorithm to perform 3D reconstruction between these camera are dif ferent due to the produced images are dif ferent as well. Kinect abilities allo ws RGB image and depth map produced, on the other hand Stereo camera has to perform another depth map acquisition algorithm by combining 2 RGB images. Numerous numbers of 3D reconstruction task can be found in capturing the site, cultural artif acts both in ground and under the sea le v els [1]. The e xtinction f actor is the most prominent issue in these area. Moreo v er , 3D imaging data also could help impro v e the accurac y of the anatomical features in order to observ e some areas before coming to the sur gery action .Furthermore, in order to perform 3D reconstruction, there are multiple approaches found in the literature re vie w such as from the broad ra ng e s of vision setup, v arious types of inputted image to construct 3D reconstruction. Thus, In this paper will describe more on those approaches. The great numbers of the resear chers along with the hardw are supports allo ws such algorithm to do high processing calculation in order to perform reconstruction task. There are some sections mentioned in part 2.. The benefits of reconstruction are to perform 3D recording, visualization, representation and reconstruction [2]. Moreo v er Tsiaf aki and Michailidou e xplained that, there are 6 benefits in performing reconstruction and visualization: limiting the destructi v e nature of e xca v ating, placing e xca v ation data into the bigger picture, limiting fragmentation of archaeological remains, cl assifying archaeological finds, limiting subjecti vity and publication delays, enriching and e xtending archaeological research. J ournal homepage: http://iaescor e .com/journals/inde x.php/IJECE Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 r 2395 Some algorithms found in the literature re vie w introduced the usage of single and multiple i mages approaches to perform 3D reconstruction. There are some characteristics of the algorithms in the literature specifically b uilt for single or multiple images, adv antages and dra wbacks e xplained in this paper . In this paper will described the vision setup by 3 cate gories as follo ws: 1. Single Camera A single camera is si mple to calibrate, computationally ef ficient more compact. Ho we v er , the y are lack of the depth information. It requires prior kno wledge from other sensor to determine the depth scale [3]. 2. Stereo Camera In stereo camera mechanism is that the images captured either using 2 equals web camera [4] or an y cameras. The y are set by a defined distance. In addition to 2 images captured, an algorithm is used to generate depth map. Ho we v er , stereo matching ha v e se v eral issue when the scene contains weekly te xtured areas, repet iti v e patterns or occlusions occur in both indoor and outdoor en vironments [5] as sho wn in Figure 1. Figure 1. Stereo Camera 3. Kinect / Structured Light / T ime of Flight Structured Light sensor is able to perform range detection, an accurate distance measurement is the output [6]. Kinect came ra is a product from Microsoft that has an RGBD camera. The product comes with nati v e SDK that allo ws user to call the API to perform some vision task such as sk eleton detection. 4. Fusion Some researchers also tried possibilities of using fusion approach where a s combining depth map pro- duced by Stereo and kinect camera to achie v e h i gher accurac y in depth map precision. T o such de v elop- ment allo ws to produce better 3D Reconstruction object, rich in features details. Range cam eras are lo w cost and ease to use t o construct 3D point clouds in r eal time. One issue arise is that the transparent and reflecti v e surf aces [7]. on the other hand, 3D model produced by stereo vision are mostly incomplete in lo w te xture re gions. The possibilities of combining both approached could lead to better depth map quality . Fusion approach is sho wn in Figure 2. Figure 2. Fusion Approach Computer vision based 3D r econstruction : A r e vie w (Ham Hanry) Evaluation Warning : The document was created with Spire.PDF for Python.
2396 r ISSN: 2088-8708 The algorithms v ary due to the characteristics of the inputted image. Therefore, in this paper we described the inputted image into 2 cate gories : single and multiple images. Single image, the characteristic image can be described as: 1. Single Still Image Single still image here using an RGB image. This image can be tak en by a re gular camera. 2. RGB-Depth Image RGB image is tak en with the setup camera that produced RGB-D format image. Mostly , the setup used is commercial camera such as Kinect, Intel real sense camera. On the other hand, the multiple images can be described as: 1. Multiperspecti v e of 2D images [8] The idea of this aprroach is to tak e some images dif ferentiate in its perspecti v e to the object. Thus the area of the object are co v ered properly using filter [9]. In addition to that, Xian-hua and Y uan-qing [10] said that in order to perform 3D reconstruction, an ef fecti v e matching of a feature is the prominent f actor in later stage. The y implemented a feature matching error elimination method based on collision detection. 2. V ideo Sequences The using of the input video sequences as kno wn as structure from motion. Sepehrinour and Kasaei e xplained that these methods are using the shared information of consecuti v e frames, in the form of tracking information of feature points in a sequence of images. The f actors may impact to the de v eloped methods: the kno wledge or lack of kno wledge of camera calibration parameters, ha ving multiple cameras with dif ferent vie wing angles or only one mo ving camera, and rigid or non-rigid shape reconstruction based on the incoming video stream. 2. T AXONOMY OF 3D RECONSTR UCTION 3D Reconstruction plays an important roles in se v eral aspects such as medical imaging data, site and cultural artif act reconstruction. (a) Medical Imaging Data Common sur gery operation procedures uses X-Ray as a reference for the doctor to operate on specific section. Ho we v er , some important features cannot be visualized well in 2D images [12]. In addition to 2D images, the accurac y may increase depends on se v eral aspects such as: number of 2D V ie ws, the image noise, and the image distortion. Magnetic resonance Images also holds an important method while considering the operation process. The gi v en output of MRI are in 2D images, ho we v er there are some literature can be found in manipulating those images into 3D space. By implementing such method, the y w ould lik e to pro v e the more features captures, the more accurate result is. a w ork from Hichem et al. introduced a geometric interpretation of the 3D model reconstruction of the blood v ess el of the human retina. Sumijan et al. [14] in their w ork introduced a method to calculate v olume Hemorrhage Brain on CT -Scan Image and 3D Reconstruction. The idea of this w ork is to calculate of the bleeding area in the brain on each image slide CT -scan. As it is said in the pre vious w ork[15], brain injury is one of the most causes that cause the death of human. In addition to the pipeline, the e xtraction the bleeding area of the brain using Otsu algorithm combining with the morphological features algorithm. Therefore by visualizing the brain v olume aim at impro ving visual enhancement for the doctor to gi v e the best medical treatment. (b) Site and cultural artif acts Reconstruction The site reconstruction has been widely an issue to the archaeology in order to capture the social, culture through the b uilding, the y do the reconstruction. Re gular camera can only allo w to capture in 2D space format. Not all the details from the b uilding can be captured and closely observ ed. Since then, by using stereo camera or Kinect mak e this task possible along with the algorithm de v eloped in the current research. The archaeological sites are not only on the ground b ut also under the sea. The reconstruction which performed under the sea rises another issue to the images captured such as de gradation quality Int J Elec & Comp Eng, V ol. 9, No. 4, August 2019 : 2394 2402 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 r 2397 if underw ater images, une v en illumination of light on the surf ace of objects. scattering and absorption ef fects [1]. (c) Nuclear Substantial Reconstruction Monterial et al. [16] used 3D image reconstruction of neutron sources that emit correlated g ammas. This aim at pre v enting nuclear threat search, safe guards and non-proliferation. This research is prominent and under supervision of le g al di vision. In addition to that, nuclear had been used as source of ener gy , yet some contro v ersies arise about the impact of harmful substantial. 2.1. Single still image appr oach The first part will describe the algorithms found in the literature re vie w using single still image. Com- pared to the multiple images, single image occurs tend to ha v e more challenges. Sax ena et al. e xplained that one of the issued is to create a depth map due to the loca l features are insuf ficient to estimate depth at a point. In addition, single still image approach is relati v ely less studied in the literature. Sax ena et al. [17] introduced a 3D depth reconstruction using a single still image. A s upervised learning approach w as proceeded by taking a training set including the unstructured indoor and outdoor en- vironments and their corresponding ground-truth depthmaps. Their proposed algorithms a w are of the global structure of the image, based on modeling depths and relationships between depths using proposed multiple spatial scales using a hierarchical, multiscale Mark o v Random Field. Ground truth were tak en using 3D scan- ner . Y an et al. [8] proposed a system called Perspecti v e T ransformer nets. The model w as b uilt by ignoring the color and te xture f actors. In addition to that, the e xperiments sho ws that e xcellent performance of the pro- posed model in reconstructing the object without ground-truth 3D v olume as supervision. The input used were pro vided by Chang et al. [18] w orks. The images input proposed is a single vie w 3D v olume reconstruction [19] with perspecti v e transformation [20] run through defined encoder -decoder netw ork that consists of a 2D con v olutional encoder , a 3D up-con v olutional decoder and a perspecti v e transformer netw orks. F an et al. [21] applied a re gion-based gro wing algorithm for 3D reconstruction by using brain MRI images. There are 3 steps in their proposed pipeline : First, the seed element is the initial state of the se gmenta- tion. Second, start the gro wing process from the seed element. There are 4 areas of the gro wth area. Ho we v er there are some defined threshold v alue to meet the pattern of gro wth. Third, us e the points which sat isfy the gro wing requirement as seed element, and continue to gro w . in addition to the result, their proposed method could achie v e 90.52% compared to Nadu [22] w orks. 2.2. RGB-depth image appr oach Zhang et al. [23] de v eloped a feature-based RGBD camera pose optimization for real-time 3D recon- struction. Their proposed w ork are ignoring corner -based feature detectors such as BRIEF and F AST due to acquired images contains huge noise around object contours. Subsequently , SURF detector w as chosen due to the f act that its rob ustness, stability , scaleable and rotation in v ariant [24]. In addition to that, SURF can be com- puted in parallel on the GPU [25]. The miss-matched pairs in feature matching can be remo v ed using RANSA C algorithm. The consistenc y of the global po s itions of matched features are track ed by proposed feature cor - respondence list and camera pose optimization both in the spatial and temporal dimension. Subsequently , in order to e v aluate the method, v ox el-hashing w as used for each camera poses compared to the proposed method. It is pro v ed that their proposed optimized camera poses outperforms the structure of the reconst ruct model for the real scene data captured by a f ast mo ving camera. Group et al. [26] e xplained that a fully con v olutional 3D denoising autoencoder neural netw ork. The y e xperimented using RGBD dataset and it is pro v ed that the netw ork could reconstruct a full sc ene from a single depth image by filling holes and hidden element. The netw ork is capable of learn the object shape by inferring similarities in geomet ry . A real-w ord dataset of table top scenes [27] w as used using KinectFusion. Their steps can be mentioned as follo ws: acquisition RGBD image using Kinect, denoising and hole filling depth channel using [28] algorithm, projection of the pix el into 3D space using preset equations, retrie v e sensor pose from accelerometer and align point cloud data, v ox elize the point cloud, and A predefined CNN layer w as trained. In addition to that, the netw ork is not constrained to a fix ed 3D shape and it is capable successfully reconstructing arbitrary scenes. Jaisw al et al. [29] used Kinect to assess 3D object modelling. The proposed pipeline are as f o l lo ws: first, 3D point cloud, a green surf ace w as placed behind and under t he object to do the histogram-based se g- Computer vision based 3D r econstruction : A r e vie w (Ham Hanry) Evaluation Warning : The document was created with Spire.PDF for Python.
2398 r ISSN: 2088-8708 mentation out the object from the RGB images. Afterw ards, RANSA C algorithm is used to perform a coarse alignment. Second, the re gistration using SIFT based [30] to o v ercome the lack structural features or under go significant changes in camera vie w . Third, global alignment is used ti eliminate inaccurac y at each re gistration that could lead to significant misalignment between the first and last frame. F ourth, 3D point cloud denoising is performed to refine the 3D object model, in this case M o ving Least Square (MLS) 3D model denoising method [31]. Fifth, surf ace reconstruction using Delaunay triangulation method [32] to con v ert 3D point clouds into meshed. Afterw ards, coloring task is performed to each v erte x and simply interpolate the color in each triangle f aces. 2.3. Multiperspecti v e of 2D images appr oach K o w alski et al. [33] created an open source system for li v e, 3D data acquisition using multiple kinect v2 Sensors. T o o v ercome the ability of the nati v e Kinect V2 SDK, the y made this fle xible frame w ork. There are 3 coordinates system of a mark ers: Kinect v2 sensor , coordinate system of a mark er which is located at a center on a gi v en mark er and the w orld coordinate. The proposed pipeline as follo ws: first calibrations were done by calibrating 2 types of defined mark ers. Subsequently the Iterati v e Closest Points (ICP) algorithm [34] w as used to refine the initial estimation. Ev angelidis et al. [5] combined lo w-resolution depth data with high resolution stereo data to o v ercome the construction of high-resolution depth maps for the range-stereo fusion problem. The input used stereo images (high resolution) and depth data (lo w resolut ion) from the range camera. The lo w resolution depth data are projected into the color data and refined a h i gh resolution sparse disparity map. Subsequently , the depth up-sampling algorithms were perform such as triangulation-based interpolation and join bilateral filter . then a re gion gro wing fusion were performed and final denser High resolution map as the result. Burns [35] introduced a te xture super resolution (TSR) method for 3D multi-vie w reconstruction. In addition, their w ork used video sequence as the input. Moreo v er to the proposed pipeline, a Photoscan from Agisoft is used to do multi-vie w stereo reconstruction and 3D mesh model. Then, optical flo w algorithm is inte grated in order to re gister each pix el of neighboring to the closest k e y-frame using KL T feature track er [36]. Afterw ards, to support rob ustness to outliers the fundamental matrix filtering of the track ed 2D points and RANSA C filtering of the 2D/3D correspondences. Due to the piece-wise af fine surf ace approximation constructed in 3D mesh, this may lead to pix els re gistration error . T o o v ercome that issue, to locate the dis- placements, an optical flo w estimation is used [37]. The object used is 2mx1m desk that has man y te xtured objects on it as gray-scale images along with the subsampling applied to it. It is acquired using a camera with 5.5mm focal length at f/2.8 mounted on a Byaer 1/18” e2v detector . There are 3 e xperiments conducted and it sho ws that the proposed methods outperforms compared to the re gistration with mesh and camera poses only , re gistration with optical flo w only . T ulsiani et al. [38] studied multi-vie w supervision for single-vie w reconstruction and a dif ferentiable ray consistenc y (DRC) term w as introduced which allo ws computing gradients of the 3D shape gi v en an ob- serv ation from an arbitraty vie w . The dataset used is called ShapeNet dataset. The follo wing steps to perform their methods are: formulation , vie w consistenc y loss function is introduced aim at measuring the inconsistenc y between a predicted 3D share and a corresponding observ ation image. shape r epr esentation , The assumption made w as it is possible to trace trays accross the v ox el grid and compute intersection with cell boundaries. The 3D shape representation is parametrized in a discretized 3D v ox el grid. Observ ation, This aim at achie ving the shape to be consistent with some a v ailable observ ation such as depth image, object fore ground mask. Also CNN model w as used as a simpl e encoder -decoder which predicts occupancies in a v ox el grid from the input RGB image. The result outperformed all the algorithms found in the literature re vie w . Martin-Brualla et al. [39] e xtended 3D time-lapse reconstruction where a virtual camera mo v es con- tinuously in time and space using internet photos. Pre vious w ork assumed a static camera, the addition of camera motion during the time-lapse produces a v ery compelling impression of parallax. The first step is a pre-processing step, computing 3D pose of the inputted image using structure from motion algorithm. Sub- sequently , the desired path has to be specified through the reconstructed scene. Then, the algorithm compute time-v arying, temporally consistent depthmaps for all output frames in the sequences. Proposed 3D time-lapse reconstruction computes t ime v arying, re gularized color profiles for 3D tracks in the scene. output video frames are reconstructed from the projected color profiles. Int J Elec & Comp Eng, V ol. 9, No. 4, August 2019 : 2394 2402 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 r 2399 2.4. V ideo sequences Sepehrinour and Kas aei [11] introduced a no v el algorithm for perspecti v e projection reconst ruction using single vie w videos of non-rigid surf aces. The system i np ut is a single vie w video that tak en in a totally natural en vironm ent. In addition to that, the features e xtracted: projecti v e depth coef ficients of all points in each of the input frames, projection matrix components (camera calibration, rotation matrix, and transmission v ector). Xu et al. [40] de v eloped underw ater 3D object reconstruction with mul tiple vie ws in video st ream via structure from motion (SFM). The y are trying to capture the inherent geometrical v ariation of 3D objects at multiple visual angles using a myring streamline A UV system with CCD camera with resolution of 480 TVL/PH and the m inimum scene illumination 0.28 lux on board. The proposed pipeline : continuous videos stream combining SFM with object tracking strate gies. An object tracking so called particle filter has been introduced in image sequence with multiple vie ws to focus on the motion trajectories of underw ater 3D objects all the time. a process of triangulation, iterati v e process, and other parameter adjustment is set for SFM algo- rithm to reco v er and estimate the pos ition of the camera calibration and the geometry of underw ater scene with sparse 3D point cloud. Lapandic et al. [41] introduced a frame w ork for automated reconstruction of 3D model from mul tiple 2D Aerial im ages using Unmanned Aerial V ehicle (U A V). The objecti v e of this w ork is to a chie v e near real-time performance with reliable accurac y and e x ecution time. The proposed pipeline as follo ws: feature detection and e xtraction using F AST algorithm and Lucas-Kanade method respecti v ely , 2D point correspondence, point cloud filtering, camera pose estimation, points triangulation and point cloud calculation. 3. DISCUSSION The oldest paper cited in this paper is 1981 and the research about 3D reconstruction is still going on. This pro v ed that the maturity of the research in this area is achie v ed. There are numerous algorithms is described in solving numerous of problem s. In addition, the commercial softw are such as Microsoft, Agisoft, intel real sense, asus and man y others companies de v elop softw are and hardw are to perform such calculation. The general pipelines found in the literature re vie ws are: first, image acquisition. There are some datasets a v ailable that can be used in order to e v aluate the per formance of the proposed algorithms. Moreo v er , chances to create o wn object using vision setup mentioned earlier in section ?? . Second, Pre-processing step by allo wing some filters applied to get the best images to construct. Third, 3D cloud points. The alignment algorithm plays an important role to get decent accurac y . Along with the refinement method in mismatched 3D cloud re gistration. F ourth, 3D reconstruction is where the te xturing and meshed are applied as the final result. 4. CONCLUSION In this paper e xplains se v eral current 3D reconstruction methods from literature re vie w . There are v arious algorithm in order to perform each step of general algorithm of 3D reconstruction. Each object con- structed required special algorithms depends on the vision setup, the te xture and size of the observ ed object. The impro v ement of the sensor could lead to the higher accurac y of creating 3D reconstruction in the future be- sides the ef ficient algorithms . Modeling using neural netw ork sho ws a great adv antages [26], [8]. The defined netw ork will try to learn the shapes and will fill the occlusion re gion automatically . A CKNO WLEDGEMENT The author also w ould lik e t o ackno wledge Bina Nusantara Uni v ersity for the grant research funding. REFERENCES [1] A. Anwer , S. S. A. Ali, and F . Meriaudeau, “Underw ater online 3D mapping and scene reconstruction using lo w cost kinect RGB-D sensor, 2016 6th International Confer ence on Intellig ent and Advanced Systems (ICIAS) , pp. 1–6, 2016. [Online]. A v ailable: http://ieee xplore.ieee.or g/document/7824132/ [2] D. Tsiaf aki and N. Michailidou, “Benefits and Problems Through the Application of 3D T echnologies in Archaeology: Recording, V isualisation, Representation and Reconstruction, SCIENTIFIC CUL TURE Tsiafaki & Mic hailidou SCIENTIFIC CUL TURE , v ol. 1, no. 3, pp. 37–45, 2015. Computer vision based 3D r econstruction : A r e vie w (Ham Hanry) Evaluation Warning : The document was created with Spire.PDF for Python.
2400 r ISSN: 2088-8708 [3] F . Santoso, M. Garratt, M. Pick ering, and M. Asikuzzaman, “3D-Mapping for V isualisation of Rigid Structures: A Re vie w and Comparati v e Study, IEEE Sensor s J ournal , v ol. PP , no. 99, pp. 1–1, 2015. [Online]. A v ailable: http://ieee xplore.ieee.or g/lpdocs/epic03/wrapper .htm?arnumber=7322186 [4] A. Harjok o, R. M. Hujja, and L. A w aludin, “Lo w-cost 3D surf ace reconstruction using Stereo camera for small object, 2017 International Confer ence on Signals and Systems (ICSigSys) , pp. 285–289, 2017. [Online]. A v ailable: http://ieee xplore.ieee.or g/document/7967057/ [5] G. D. Ev angeli dis, M. Hansard, and R. Horaud, “Fusion of Range and Stereo Dat a for High-Resolution Scene-Modeling, IEEE T r ansactions on P attern Analysis and Mac hine Intellig ence , v ol. 37, no. 11, pp. 2178–2192, 2015. [6] G.-v . J. M and M.-v . J. C, “Simple and lo w cost scanner 3D system based on a T ime-of-Flight ranging sensor, pp. 3–7, 2017. [7] R. Ra v anelli, A. Nascetti, and M. Crespi, “Kinect V2 and Rgb Stereo Cameras Inte gration for Depth Map Enhancement, ISPRS - International Ar c hives of the Photo gr ammetry , Remote Sensing and Spatial Information Sciences , v ol. XLI-B5, no. July , pp. 699–702, 2016. [Online]. A v ailable: http://www .int-arch- photogramm-remote-sens-spatial-inf-sci.net/XLI-B5/699/2016/isprs-archi v es-XLI-B5-699-2016.pdf [8] X. Y an, J. Y ang, E. Y umer , Y . Guo, and H. Lee, “Perspecti v e T ransformer Nets: Learning Single-V ie w 3D Object Reconstruction without 3D Supervision. [9] Q. Hao, R. Cai, Z. Li, L. Zhang, Y . P ang, F . W u, and Y . Rui, “Ef ficient 2D-to-3D correspondence filtering for scalable 3D object recognition, Pr oceedings of the IEEE Computer Society Confer ence on Computer V ision and P attern Reco gnition , no. 1, pp. 899–906, 2013. [10] J. Xian-hua and Z. Y uan-qing, “Error Elimination Algorithm in 3D Image Reconstruction, v ol. 12, no. 4, pp. 2690–2696, 2014. [11] M. Sepehrinour and S. Kasaei, “Perspecti v e reconstruction of non-rigid surf aces from single-vie w videos, 2017 25th Ir anian Confer ence on Electrical Engineering , ICEE 2017 , no. Icee20 17, pp. 1452–1458, 2017. [12] J. Y ao and R. T aylor , “Assessing accurac y f actors in deformable 2D/3D medical image re gistration using a statistical pelvis model, Pr oceedings of the IEEE International Confer ence on Computer V ision , v ol. 2, no. Iccv , pp. 1329–1334, 2003. [Online]. A v ailable: http://www .scopus.com/inw ard/record.url?eid=2- s2.0-0344983014&partnerID=tZOtx3y1 [13] G. Hichem, F . Chouchene, and H. Belmabrouk, “3D model reconstruction of blood v essels in the retina with tub ular structure, International J ournal on Electrical Engineering and Informatics , v ol. 7, no. 4, pp. 724–734, 2015. [14] S. Sumijan, S. Madenda, J. Harlan, and E. P . W ibo w o, “Hybrids Otsu method, Feature re gion and Mathematical Morphology for Calculating V olume Hemorrhage Brain on CT -Scan Image and 3D Reconstruction, TELK OMNIKA (T e lecommunication Computing Electr onics and Contr ol) , v ol. 15, no. 1, p. 283, 2017. [Online]. A v ailable: http://journal.uad.ac.id/inde x.php/TELK OMNIKA/article/vie w/3146 [15] F . Care gi v er , A. Introduction, D. T raumatic, M. Tbi, M. Tbi, S. Tbis, A. Tbi, T . B. I. Penetration, F . V io- lence, C. Changes, and P . Changes, “F act Sheet T raumatic Brain Injury , pp. 1–6, 2018. [16] M. Monterial, P . Marleau, and S. A. Pozzi, “Single-V ie w 3-D Reconstruction of Correlated Gamma- Neutron Sources, IEEE T r ansactions on Nuclear Science , v ol. 64, no. 7, pp. 1840–1845, 2017. [17] A. Sax ena, S. H.Chung, and A. Y . Ng, “Depth reconstruction from a single still image.” Ijcv , v ol. 74, no. 1, 2007. [18] A. X. Chang, T . Funkhouser , L. Guibas, P . Hanrahan, Q. Huang, Z. Li, S. Sa v arese, M. Sa vv a, S. Song, H. Su, J. Xiao, L. Y i, and F . Y u, “ShapeNet: An Information-Rich 3D Model Repository, 2015. [Online]. A v ailable: http://arxi v .or g/abs/1512.03012 [19] D. J. Rezende, S. M. A. Eslami, S. Mohamed, P . Battaglia, M. Jaderber g, and N. Heess, “Unsupervised Learning of 3D Structure from Images, 2016. [Online]. A v ailable: http://arxi v .or g/abs/1607.00662 [20] J. W u, T . Xue, J. J. Lim, Y . T ian, J. B. T enenbaum, A. T orralba, and W . T . F reeman, “Single image 3D interpreter netw ork, Lectur e Notes in Computer Science (including subseries Lectur e Notes in Artificial Intellig ence and Lectur e Notes in Bioinformatics) , v ol. 9910 LNCS, pp. 365–382, 2016. [21] B. F an, Y . Rao, W . Liu, and Q. W ang, “Re gion-Based Gro wing Algorithm for 3D Reconstruction from MRI Images, pp. 521–525, 2017. [22] T . Nadu, “Brain T umor Se gmentation of MRI Brain Images through FCM clustering and Seeded Re gion Gro wing T echnique, v ol. 10, no. 76, pp. 427–432, 2015. Int J Elec & Comp Eng, V ol. 9, No. 4, August 2019 : 2394 2402 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 r 2401 [23] M. Zhang, Z. Zhang, and W . Li, “3D Model Reconstruction based on Plantar Image s Feature Se gmen- tation, pp. 1–5, 2017. [24] L. Juan and O. Gwun, “A comparison of sift, pca-sift and surf, International J ournal of Ima g e Pr ocessing (IJIP) , v ol. 3, no. 4, pp. 143–152, 2009. [25] W . Y an, X. Shi, X. Y an, and L. W ang, “Computing OpenSURF on OpenCL and general purpose GPU, International J ournal of Advanced Robotic Systems , v ol. 10, pp. 1–12, 2013. [26] M. L. Group, M. Intel, D. Ireland, A. P alla , D. Molone y , and L. F anucci, “Fully Con v olutional Denoising Autoencoder for 3D Scene Reconstruction from a single depth image, no. Icsai, pp. 566–575, 2017. [27] M. Firman, O. M. Aodha, S. Julier , and G. J. Brosto w , “Structured Prediction of Unobserv ed V ox els from a Single Depth Image, 2016 IEEE Confer ence on Computer V ision and P atte rn Reco gnition (CVPR) , pp. 5431–5440, 2016. [Online]. A v ailable: http://ieee xplore.ieee.or g/document/7780955/ [28] S. Liu, C. Chen, and N. K ehtarna v az, “A computationally ef ficient denoising and hole-filling method for depth image enhancement, v ol. 9897, p. 98970V , 2016. [Online]. A v ailable: http://proceedings.spiedigitallibrary .or g/proceeding.aspx?doi=10.1117/12.2230495 [29] M. Jaisw al, J. Xie, and M. T . Sun, “3D object modeling with a Kinect camera, 2014 Asia-P acific Signal and Information Pr ocessing Association Annual Summit and Confer ence , APSIP A 2014 , 2014. [30] J. Xie, Y . Hsu, R. Feris, and M. Sun, “Fine re gistration of 3D point clouds with iterati v e closest point using an RGB-D camera, Cir cuits and Systems (ISCAS . . . , pp. 1–4, 2013. [Online]. A v ailable: http://staf f.w ashington.edu/junx/publication/Fine Re gistra- tion ISCAS13.pdf%5Cnhttp://ieee xplore.ieee.or g/xpls/abs all.jsp?arnumber=6572486 [31] H. A vron, A. Sharf, C. Greif, and D. Cohen-Or , < sub > 1 < /sub > -Sparse reconstruction of sharp point set surf aces, A CM T r ansactions on Gr aphics , v ol. 29, no. 5, pp. 1–12, 2010. [Online]. A v ailable: http://portal.acm.or g/citation.cfm?doid=1857907.1857911 [32] M. Isenb ur g, Y . Liu, J. She wchuk, and J. Snoe yink, “Streaming computation of Delaunay triangulations, A CM T r ansactions on Gr aphics , v ol. 25, no. 3, p. 1049, 2006. [Online]. A v ailable: http://portal.acm.or g/citation.cfm?doid=1141911.1141992 [33] M. K o w alski, J. Naruniec, and M. Daniluk, “Li v e Scan3D: A F ast and Ine xpensi v e 3D Data Acquisition System for Multiple Kinect v2 Sensors, Pr oceedings - 2015 Int ernational Confer ence on 3D V ision, 3D V 2015 , pp. 318–325, 2015. [34] P . Besl and N. McKay , “A Method for Re gistration of 3-D Shapes, pp. 239–256, 1992. [35] C. Burns, “T e xture Super -Resolution for 3D Reconstruction, pp. 4–7, 2017. [36] J.-y . Bouguet, V . T arasenk o, B. D. Lucas, and T . Kanade, “Pyramidal Implementation of the Lucas Kanade Feature T rack er Description of the algorithm, Ima ging , v ol. 130, no. x, pp. 1–9, 1981. [37] A. Plyer , G. Le Besnerais, and F . Champagnat, “Massi v ely parallel Lucas Kanade optical flo w for real- time video processing applications, J ournal of Real-T ime Ima g e Pr ocessing , v ol. 11, no. 4, pp. 713–730, 2016. [38] S. T ulsiani, T . Zhou, A. A. Efros, and J. Malik, “Multi-vie w supervision for single-vie w reconstruction via dif ferentiable ray consistenc y, Pr oceedings - 30th IEEE Confer ence on Computer V ision and P attern Reco gnition, CVPR 2017 , v ol. 2017-Janua, pp. 209–217, 2017. [39] R. Martin-Brualla, D. Gallup, and S. M. Seitz, “3D T ime-Lapse Reconstruction from Internet Photos, International J ournal of Computer V ision , v ol. 125, no. 1-3, pp. 52–64, 2017. [40] X. Xu, R. Che, R . Nian, and B. He, “Underw ater 3D Object Reconstruction with Multiple V ie ws in V ideo Stream via Structure from Motion, pp. 0–4, 2016. [41] D. Lapandic, J. V elagic, and H. Balta, “Frame w ork for automated reconstruction of 3D model from mul- tiple 2D aerial images, Pr oceedings Elmar - International Symposium Electr onics in Marine , v ol. 2017- Septe, no. September , pp. 18–20, 2017. Computer vision based 3D r econstruction : A r e vie w (Ham Hanry) Evaluation Warning : The document was created with Spire.PDF for Python.
2402 r ISSN: 2088-8708 BIOGRAPHY OF A UTHORS Hanry Ham is a lecturer and research assistant at Bina Nusantara Uni v ersity with Mas- ter of Engineering from The Sirindhorn International Thai-German Graduate School of Engineering in Thailand and German (2016). He obtained Bachelor De gree in Computer Science from Bina Nusantara Uni v ersity (Indonesia) in 2014. His researches are in fields of image processing, computer vision and computer graphics. He is af filiated with IEEE as student member . Besides, he is also in v olv ed in student associations, and committee of se v eral competitions such as BNPCHS and A CM-ICPC Re gional Asia Site. J ulian W esley is a lecturer at Bina Nusantara Uni v ersity with Ma ster of Computer Sci- ence (M.TI.) major from Bina Nusantara Uni v ersity in 2016. His researches are in fields of image processing, computer vision, and virtual reality . Besides, he is also w ork as a technology consultant who focused on IT financial industries. He is leading a R&D team in Emerio Indonesia and guide intern students from multiple uni v ersities in Indonesia. Hendra is a lecturer at Bina Nusantara Uni v ersity . He w as born in T anjungpandan, 18 July 1992. He completed his bachelor de gree in Bina Nusantara Uni v ersity on 2010. Subsequently he obtained his master de gree on 2018. Both de gree are in Information T echnology . No w , he is w orking as a Softw are Engineer at a start-up compan y in Indone- sia. Int J Elec & Comp Eng, V ol. 9, No. 4, August 2019 : 2394 2402 Evaluation Warning : The document was created with Spire.PDF for Python.