A Literature Review on Vehicle Detection and Tracking in Aerial Image Sequence using Deep Learning T

International Research Journal of Engineering and Technology (IRJET) e ISSN:2395 0056

Volume: 09 Issue: 06 | June 2022 www.irjet.net p ISSN:2395 0072

A Literature Review on Vehicle Detection and Tracking in Aerial Image Sequence using Deep Learning Technique

Abstract: Visual multi object tracking that is both robust and high performing is indeeda key difficulty in computer vision, particularly with in context of drones. Small target recognition and tracking in UAV situations is problematic for standard Multi Object Tracking (MOT) techniques based on the tracking by detection paradigm. We will be performing real time vehicle detection and tracking on Aerial Image Sequences using different AIML approach which comprehensively includes techniques such as Image Processing, Pattern Recognition and Computer Vision. It has a wide variety of applications encompassing Visual Surveillance, Traffic Control, Digital Forensics and Human Computer Interaction.

I. INTRODUCTION

Visuallymulti objectmonitoringthatisbothresilient & high performing is indeed a major difficulty in computer vision(CV), particularly in the context of UAV. Also with hugepopularisation for commercialized unmanned aerial vehicles (UAVs) as well as the advancement of OpenCV&AIMLtechnologies,dronedetectionmethodsare becoming a hot topic forresearchers. Auto navigation, campus security surveillance, & catastrophe assistance have all become easier thank to effectivevideo image computationaltechniques and powerful deep neural networks.

II. SURVEY

[1]MultipleObjectTrackingPerformance: The CLEARMOT Metrics

Metricsthatdescribe,hequalityandkeycharacteristics in numerous object tracking systems must be studied and compared in accordanceto carefully analyse and evaluate their performance. Regrettably, there has yet to be agreement on such a range of generally validmeasures. They present two new measures for evaluating MOT systems in this paper. Multiple object tracking precision (MOTP) as well as multiple object tracking accuracy (MOTA) are suggested benchmarks that can be used for a variety of monitoring activities & permit for objective

contrastof tracking systems' primary features, like accuracy atlocating targets, precisionatrecognising targetconfigurations, but also way to detect targets on consistentbases.Theyputtheproposedmetricstothetest inaseriesofglobalevaluationworkshopstoseehowuseful and expressive they were. The CLEAR workshops in 2006 and 2007 featured a wide rangeof monitoring activities whereby a big number of modelswere tested & evaluated. Theirstudiesfindings revealthat itssuggested measures accurately reflect the numerous methods' qualitiesand shortcomings in a simple and directmanner, helps in easy evaluation in performance, thusrelevant toward a wide rangeofcircumstances.

[2] Fully Convolutional Siamese Networks for Object Tracking

Traditionally, the problem of arbitrary target tracking was tackled by developing a system of the targets arrival entirelyonline,withonlythevideoastrainingdata.Despite their effectiveness, these approaches' online only methodology limits using depthinformation whichcan be studied. Many efforts have actually been developed towards harnessingdeep convolutional networks' descriptive ability. Once the target to monitor isn't determined ahead of time, Stochastic Gradient Descent online is required inadjustingthe network's parameters, risking overall system's speed. For object detection in video, a basic tracking method is combined with a novel fully convolutional Siamese network that has been trained end to end on the ILSVRC15 dataset. The tracker reaches state of art success in various tests with the minimal brevity.Itworksatfpsthatarefasterthanactual.

[3]SimpleOnlineAndRealtimeTracking

This research looks at a realistic approach for monitoring many items, with the primary objective of associating objects successfully for online and real time operations. The study claims that recognition abilityis a criticalcomponentindeterminingdetectionaccuracy,with modifyingitsdetectorboostingtrackingefficiencybyupas 18.9 percentage. In contrast to many batch based tracking systems, this research focuses on online tracking, where

International Research Journal of Engineering and Technology (IRJET) e ISSN:2395 0056

Volume: 09 Issue: 06 | June 2022 www.irjet.net p ISSN:2395 0072

thetrackerisonlyshowndetectionsfromthepreviousand current frames. Despite just employing a simple mix of existing techniques such as the Kalman Filter and the Hungarian algorithm for the tracking components, this approach achieves tracking accuracy similar to state of the art online trackers. This research looks at a realistic approach for monitoring many items, with the primary objective of associating objects successfully for online and real time applications. The study claims that detection quality is a critical factor in determining tracking performance, withchanging thedetector boostingtracking performancebyupto18.9%. Thetrackeralsoupdatesata rate of 260 Hz, which is nearly 20 times faster than other state of the art trackers due to the simplicity of the trackingmethod.

[4] High Speed Tracking by Detection without Using ImageInformation

As the performance in object detectors increases, the foundation for a tracker becomes significantly more trustworthy. The problems for a successful tracker have changed as a result of this, as well as the increased use of higher frame rates. As a result of this shift, considerably simpler tracking algorithms may now compete with more complex systems for a fraction of the processing cost. This paperoutlinesandillustratessuchamethodbyconducting extensive tests with a range of object detectors. The proposed technique can easily operate at 100K fps on the DETRAC vehicle tracking dataset, beating the state of the art. The notion of a passive detection filter is used to analyse a very simple tracking technique in this research. Due to its modest computing footprint, the suggested approach can serve as a basic predictive model for other trackers and provide an appraisal of the necessity of additional efforts in the tracking algorithm. It also permits reviewing tracking benchmarks to evaluate if the specific concerns they indicate (for instance, missed detections, frame rate, etc.) are within the capabilities of existing algorithms.

[5] Cascade R CNN: Delving into High Quality Object Detection

The this study, they presented the Cascade R CNN, a multi stage object recognition framework for developing high quality object detectors. Overfitting during learning and quality disparity during inference have both been demonstrated to beavoided using thisdesign. Onthehard COCO and famous PASCAL VOC datasets, the Cascade R substantial CNN's and consistent recognition improvements show that effective object detection necessitates modelling and knowledge of several corroborating aspects. The Cascade RCNN has been shown toworkwithawiderangeofobjectdetectionarchitectures. They hope it will be useful in a variety of future object detectionresearchprojects.

[6]Real Time ‘Actor Critic’ Tracking

Object tracking is one of the most important step in object detection and tracking. In this paper Actor Critic framework been used, where the 'Actor' model seeks to infer the best option in a continuous action space, causing the tracker to move the bounding box to the object's current location. The 'Critic' model is used for offline training to create 'Actor Critic' framework along with reinforcement learning as well as a Q value for directing both the 'Actor' and 'Critic' deep network learning processes. Visual tracking is viewed as a dynamic search processinwhichthe'Actor'modeloutputsonlyoneaction to locate the tracked object in each frame. Offline training of better policy for finding the best result is done using reinforcement learning. Furthermore, the 'Critic' network serves as a verification system for both offline and online instruction. Using popular benchmarks, The suggested tracker gets contrasted to certain state of art trackers, as well asthe stimulation finding revealthat itperforms well inactual.

[7]HybridTaskCascadeforInstanceSegmentation

At instance level, segmentation process is basic computer vision job which identifies objects per pixel. In real world settings like automated driving and video surveillance, precise and reliable feature extraction is challenging to achieve. Cascade would be a basic but efficient design which has increased results on a wide range of workloads. A basic Cascade R CNN and Mask R CNNcombinationproducesjustalittleboost. Thesecretto good instance segmentation cascade would be to completelyuseinverseinteractionacrossdetectionaswell as segmentation so that to discover a more effective technique.

TheyproposeHybridTaskCascade(HTC)inthispaper, that is different in two key ways: (i) it intertwines these twotasksforsimultaneousmulti stagecomputation,rather thanconductingcascadedrefinementonthemindividually; and (ii) employs a convolutional section in order to give spatial features, that help distinguish difficult frames in cluttered background. Bounding box analysis plus masked predictions is coupled inside a multi tasking way at each step of HTC. At certain stages, easily applicable within its maskedsectionsarealsogiven themaskedcharacteristics in every step are combined and supplied to the next. The whole design increases data flow inside the activities as well as stages, resulting in good refined predictions at all levels and more reliable forecasts overall. HTC is simple insettingup&maybeprogrammedfrombeginningtoend. Itgained2.6%&1.4%greatermaskedAPthanthatofthe Masked R CNN as well as Cascade Masked R CNN benchmarks,respectively,ondifficultCOCObenchmark

International Research Journal of Engineering and Technology (IRJET) e ISSN:2395 0056

Volume: 09 Issue: 06 | June 2022 www.irjet.net p ISSN:2395 0072

[8]MMDetection:Open MMLab Detection Toolbox and Benchmark

CVtasksliketargetrecognition&instancesegmentation are both fundamental. Also the detection framework's pipeline is typically more complicated than that of classificationjobs,andvariousimplementationparameters might produce drastically different results. In this paper they are having the goal of providing a high quality codebase and unified benchmark, for this they have built MMDetection model. Major features of MMDetection (1) Modular design They split the detection framework into separate components, allowing users to quickly build a customised object detection framework by combining different modules. (2) Out of the box support for multiple frameworks. Popular and current detection frameworks are supported by the toolbox. (3) High efficiency GPUs handle all fundamental bbox and mask operations. Other codebases, such as Detectron, maskrcnn benchmark, and SimpleDet, have training speeds that are quicker or equivalent. (4) State of the art The toolbox is based on the software created by the MMDet team, who won the 2018COCODetectionChallenge.

[9] A Cost Effective Tracking System for Small UnmannedAerialSystems

The paper describes a object tracking system for UAVsystems is described. It is built on object recognition with a sum of absolute differences similarity measure. Every algorithm loop contains a template change that incorporates new data so that the former stays valid, allowing the track to proceed even if the object's appearancechanges.Italsoenablesthehumanoperatorto entertargetsmanuallyandreceivedatafromotherpicture processingunits..

Thepapersupportstheimportanceofuavvisionmodel suggesting a small, imprecise, non gyro stabilized architectural method for detecting & monitoring many targets in real time by not making any rigid assertions on trajectories, depending onto limited input of their surroundings and also with avoiding necessity motion compensation, which usually necessitates the need for a inertialmeasurementunit(IMU). Thisapproacheliminates the additional expense & cost of a binocular model and thoseofgyro stabilized turret instead opting for small, unstablesystem.Nnotionthatcameradoesn'talwaysneed to be corrected for such modelto work reduces the building & configuration procedure even further. Even more simplifies the assembling and setup procedure. The ideacanruninrealtimeonlow cost,low powercomputer systems thanks to its ability to track several things at a cheap computational cost. It's highly crucial because automated driving vehicleswith restricted payload capacity, whom it is impractical to have a substantial

energybackupforsustainingthecomputer.Despitethefact that current advances in recent computer architecture & powerful batteries assurances that meettheissue by supplying more productive, advancedmachineriesand greaterpowerintensitybattery,aurgeofasimplerstyleto vision systems integration will almost definitely stay important. This is particularly true for smaller unmanned aerialsystems,suchastheMaxxiJoker2systememployed inthisresearch.

[10] The research onvisual object tracking

basedon an adaptivecombinationkernel

algorithm

The research suggested a visual object tracking approach based on the Adaptive Combination Kernel to increase the resilience to intricate transitions of multiple objects and a complex backdrop picture. The object tracking approach has indeed been divided down into separate subtasks to approximate the object's details: Translation Filter as well as Scale Filter. At commence, the Translation Kernel Tracker utilizes a new Linear Kernel Filteraswell asGaussianKernel Filter(GKF) pair.The goal function, which incorporates not just overfitting problem but also the highest amount of response output for every kernel, was used to determine the weight coefficients in boththeLinearKernelfilteraswellasGKF.Theadvantages of both the local plusglobal kernels are combined in the Adaptive Combination Kernel. Next, the tracker position was identified by using response result of the dynamic combination kernel correlation filter. Furthermore, the interpretation filter has been constructed with a scene adaptive training data depending on the highestresponse score. The effective learning speed can be used to update the translating filters. Lastly, the item scale was estimated using a 1Dmagnitude filters. Compared to previous techniques, the proposed algorithm is more robust to deformationandocclusion.

[11] Dual Channel Convolutional Neural Networks BasedImageSuper ResolutionAlgorithm

By using picture super resolution technique on single network, it's really challenging in accomplishing combinedhigh quality pattern reconstruction and quick converging. To overcome the drawbacks of earlier approaches, the current study proposes a picture high resolution strategy built upon dual channel CNN (DCCNN). A deep tunnel as well as a shallow tunnel were created in the system model's novel architecture. The shallow tunnelwas largely employed to maintain the original picture's general shape, whereas the deep tunnel has been used to retrieve specific featuredata. To begin, during featureextractionphase,theleftoverframewasaltered,as well as the channel's nonlinear mapping capability was increased. After the characteristic mapping scale was reduced,thepicture'seffectivefeatureswererecovered.

International Research Journal of Engineering and Technology (IRJET) e ISSN:2395 0056

Volume: 09 Issue: 06 | June 2022 www.irjet.net p ISSN:2395 0072

During up sampling step, deconvolutional kernel's general variables are tweaked, and high bandwidthnetwork degradation were minimised. High resolution texturearea may get recreated iteratively employing long as well as the small data chunksin recreation steps, increasing texture information recovery even further. Second, the convolutional kernel of narrow network was adjusted for decreasing the variables, Suring that the whole contour of picture is recovered & the channel narrowed faster Lastly, double channel error rateis simultaneously changed to increase capability to fit featuresintermsofachievingamostrecenthigh resolution pictureresult.

[12] Enhanced Hierarchical Principal Component AnalysisMethodfor DetectingSaliency

A saliency identification technique formedupon Hierarchical Principal Component Analysis (HPCA) was created with in study to solve the challenges of existing important item identificationapproaches, such as severe environmentalnoise, lessaccuracy, and high processing performance. After transforming the RGB picture to monochrome, the monochromeimagewas split in eight layers and used the digital surface stratification approach. Important subject data correlates to the layered image characteristics in each picture layer. Second, the monochromepictureis reinitialisedusing the monochromecolour conversion technique, which uses the initial image's colour layout like a source images, resulting in a tiered picture that not only represents the initial structuralcharacteristicsandalsoeffecientlyconservesthe initial picture's color information. Moreover, Principle Component Analysis (PCA) was used on layered picture to identify the structural and colour differential features of everytierwithinprinciplecomponentorientation.

To get a saliency layouthaving greatresilience as well astoimproveourfindingsfurther,twocharacteristicshave been combined: recognized prviously wereadded to picture organisation, which may pinpoint its photograph's topic around the image's centre. Finally, the entropy computation was used to produce the idealized picture as from multilayered saliency layout; the best layoutcontains the lowest background as well as most clearly saliency entitiescomparedtotheothers.

[13] Online Multi Object Tracking Using CNN based Single Object Tracker with Spatial-Temporal Attention Mechanism

ACNN basedarchitectureforonlineMOTisproposedin thisresearch.Thisframeworkmakesuseoftheadvantages of single object trackers When it comes to altering the shape of models & searching for the objective within next frame, There are concerns with processing efficiency and occlusion induced drifting outcomes when employing the

single object tracker to MOT. The given approach achieves computational efficiency besides wanting to share characteristics and using ROI Pooling forobtainingindividual characteristicto every target. The appearance model is adjusted in every target using any target specific CNN layers learned online. Introduction ofthespatial temporalattentionmechanism(STAM)inthe framework was carried outto deal with drift caused by occlusion and target engagement. A target's visibility plan waslearntandusedtoinferitsspatialattentionplan.

Characteristics are then weighted using the spatial attention map. Furthermore, the occlusion state can be evaluated using the visibility mapping, that regulates a continuous updating mechanism using weighted loss upon training samples having varying occlusion states over multipleframes. It's possible to think of it as a temporal attention process. On the rigorous MOT15 and MOT16 benchmark datasets, the suggested approach obtains 34.3 percentand46.0percentinMOTA,respectively.

[14]DeepLearningin VideoMulti ObjectTracking

Although the target's presence is understood in advancein SOT, MOT needs the detection stage for detectingthe objects which might leave but rather reenter theframe. Oneofmost difficultaspectofmonitoringmany objects at once is the numerous occlusions as well as interactions amongst target, which might sometimes seem toward being identical. Tracking by detection is the standard strategy used in MOT algorithms: The tracking procedure is guided by a collection of detections (i.e. enclosing boxdenotingallobjectsinsidepicture)extracted outfromimagesequences.

The following steps are found in the great majority of MOTalgorithms:

• Object detection stage: The targetidentification methodanalyses every input videoframe using bounding boxes to discover targetbelonging here to target class, oftencalledby'detections'asindomainofMOT..

• Extraction of features prediction stage: One or many feature retrievingtechniques investigate these identificationplustracklets in order to obtain information about look, movement, andinteractivity. Each monitored target's upcoming location may be predicted using a motionprediction.

• Affinity stage: By giving the same ID for detections which indicate the same target, similaritymetrics are utilised to correlate detections with trackletspertainingtotarget.

• Association stage: By giving the same ID for identificationwhich indicate the same object,

International Research Journal of Engineering and Technology (IRJET) e ISSN:2395 0056

Volume: 09 Issue: 06 | June 2022 www.irjet.net p ISSN:2395 0072

similaritymetrics have been used to correlate identificationwithtrackletspertainingtothetarget.

[15]JointMulti view Face Alignmentin theWild

In this paper, Using face detection model followed with a deformable model fitting on the bounding box is the de facto approach to estimatefacial landmarks. It entails two majorissues:



A detection & deformable fitting processes were done separately, as well as the detector may or may not offer the optimum initialization tofitting step,

that is faster andprecise but also discovers many items in the image; about 95% among all targetsinframecan be recognised, yet every frame takes on average 140 milliseconds. In general, their experiments show that SSDoutperformstheHDDmodelinthelongterm.

[17] Rich feature hierarchies for accurate object detection andsemanticsegmentation



A look of a face varies greatly with postures, makingchangeablefacialfittingextremelydifficult, necessitatingtheusageofseveralimages

They demonstrate the first joint. To the best of their knowledge, multiview convolutional networks can handle significantposturalvariancesacrossfacialwithinfieldand neatly integrate facialidentification and facial landmark localisation tasks. Currently present combined facialdetection & landmark positioning methods mainly evaluatelimitednumberoflandmarks.Theirtechniquecan recogniseaswellasmatchahugeproportionofmarkersin mid frontal (68 landmarks) as well as profiles (39 landmarks) faces. They put their system through its paces onrangeofsamples,namelyCOFW,300W,IBUG,aswellas thecurrent Menpo Standard both for mid frontal as well asprofile faces. There isconsiderable enhancementover state of art methods in deformable face recognition on 300VW benchmark. Also on FDDB and MALF benchmark, theyalsogivesstateoftheartfacerecognitionresults.

[16] Object Detection from Video Taken by Drone throughCNN

The study used CNN to recognise objects in footage taken by the drone. They suggest employing CNNs in this work to allow drones can discriminate between different entitytypeslikebuildings,automobiles,trees,andhumans. Despite the fact that convolutional neural networks are computationally expensive, they are trained with smaller picturedatasetsusingtheapproachtransferlearning.They used TensorFlow's sophisticated object detection API in their project, which allowed them to quickly build a new model and deploy it for detection. According to the data, the two models' detection rate for houses, trees, cars, & pedestrians is fairly good, with just an mean of more than 85%andapeakof99percent.Theycomparedthememory consumption, speed, and accuracy of two current object detectors in an experiment. In comparison to Faster RCNN models, SSD concepts give more importance to size,ratio, aswellaspredictedsamplelocation,withanaverageframe time of 115 ms and a low target identification rate. R CNN

According to this paper, Object detection performance has been stagnant in previous years. Complex ensembles combine several low level picture features with high level context from object detectors and scene classifiers to produce the best results. This research proposes a simple and scalable object detection technique that improves on the best recent findings on PASCAL VOC 2012 by 30 percent. They were able to attain this level of success because to two key observations. This one was to localise and segment objects using bottom up region recommendationsutilisinghigh capacityCNN.The2ndisa strategyofteachinglargeCNNswhenlabelledtrainingdata islimited.Theydemonstratethatpre trainingthenetwork for an auxiliary task with a large amountofinformation(objectclassification) and afterwards tuning the system for primary goal with minimal input is very effective (detection). They believe that the "guided pre training/domain specific finetuning" approach will work well for a range of vision difficulties with little data. They end by emphasising how remarkable it is that they were able to obtain these results utilising a combination of traditional computer vision(CV) as well as deeplearning(DL)tools.Insteadbeingantagonisticpathsof rationalresearch,theyarecloselyintertwined.

[18] DeepResidualLearningforImageRecognition

According to paper, hierarchical neural networks aredifficult to be trained. They offer an residual learning strategy for significantly deeper training modelsthan earlier used models. Researchers specifically reformulate thelevelsastraining residual methodswith relation to the levelsinputs, rather than learning unreferenced functions. Researchers presented significant experimental proof suggestingthat residual networks is simpler to implement and that increasing complexity can boost performance. Authors used the ImageNet datato evaluate residual nets having put to 152 levels of complexity, that are 8 layers deeper to VGG networks whereashave less intricacy. The aggregation of such residue nets scores 3.57 % error just on ImageNet test set. Thework won 1st spotwith in ILSVRC 2015 classification problem. CIFAR 10 study using 100 1000levelsisshown.Thecomplexityindepictionsis crucialformanyimageidentificationtasks.Onlybecauseof their exceptionally deep depictioncan they accomplish an 28 % notable improvement onCOCO object recognition samples.TheircontributionstowardsILSVRC&COCO2015

International Research Journal of Engineering and Technology (IRJET) e ISSN:2395 0056

Volume: 09 Issue: 06 | June 2022 www.irjet.net p ISSN:2395 0072

contestsuseddeepresidualnetworksasthebasis,andthus won 1st spotin the ImageNet identification, ImageNet localisation, COCO recognition, as well as inCOCO segmentationtasks.

REFERENCES

1. Bernardin K, Stiefelhagen R (2008) Evaluating multiple object tracking performance: the CLEAR MOT metrics. EURASIPJournal onImage andVideo Processing, 2008, 1 10

2. Bertinetto L, Valmadre J, Henriques JF, Vedaldi A, Torr PH (2016) Fully convolutional siamese networks for object tracking. In European conference on computer vision(pp.850 865).Springer,Cham

3.BewleyA,GeZ,OttL,RamosF,UpcroftB(2016)Simple onlineandrealtimetracking.In2016IEEEInternational Conference on Image Processing (ICIP) (pp. 3464 3468).IEEE

4. Bochinski E, Eiselein V, Sikora T (2017) High speed tracking by detection without using image information. In 2017 14th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS) (pp.1 6).IEEE

5.CaiZ,VasconcelosN(2018)CascadeR CNN:Delvinginto highqualityobjectdetection.InProceedingsoftheIEEE conference on computer vision and pattern recognition (pp.6154 6162)

6. Chen B, Wang D, Li P, Wang S, Lu H (2018) Real time'Actor Critic'Tracking. In Proceedings of the European Conference on Computer Vision (ECCV) (pp. 318 334)

7. Chen K, Pang J, Wang J, Xiong Y, Li X, Sun S ... Loy CC (2019) Hybrid task cascade for instance segmentation. In Proceedings of the IEEE conference on computer visionandpatternrecognition(pp.4974 4983)

8. Chen K, Wang J, Pang J, Cao Y, Xiong Y, Li X, ... Zhang Z (2019) MMDetection: Open mmlab detection toolbox andbenchmark.arXivpreprintarXiv:1906.07155

9. Chen Y, Wang J, Liu S, Chen X, Xiong J, Xie J, Yang K (2019)Multiscale fast correlation filtering tracking algorithmbasedonafeaturefusionmodel.Concurrency andComputation:PracticeandExperience,e5533

10. Chen Y, Wang J, Xia R, Zhang Q, Cao Z, Yang K (2019) The visual object tracking algorithm research based on adaptive combination kernel. J Ambient Intell HumanizedComput10(12):4855 4867

11. Chen Y, Wang J, Chen X, Sangaiah AK, Yang K, Cao Z (2019) Image super resolution algorithm based on dual channel convolutional neural networks. Appl Sci 9(11):2316

12. Chen Y, Tao J, Zhang Q, Yang K, Chen X, Xiong J, ... Xie J (2020)SaliencyDetectionviatheImprovedHierarchical Principal Component Analysis Method. Wireless CommunicationsandMobileComputing,2020

13. Chu Q, Ouyang W, Li H, Wang X, Liu B, Yu N (2017) Online multi object tracking using CNN based single object tracker with spatial temporal attention mechanism. In Proceedings of the IEEE International ConferenceonComputerVision(pp.4836 4845)

14.CiaparroneG,SánchezFL,TabikS,TroianoL,Tagliaferri R,HerreraF(2020)Deeplearninginvideo multi object tracking:Asurvey.Neurocomputing381:61 88

15. Deng J, Trigeorgis G, Zhou Y, Zafeiriou S (2019) Joint multi view face alignment in the wild. IEEE TransactionsonImageProcessing28(7):3636 3648

16. Fan, D. P., Wang, W., Cheng, M. M., & Shen, J. (2019). Shiftingmoreattentiontovideosalientobjectdetection. In Proceedings of the IEEE conference on computer visionandpatternrecognition(pp.8554 8564).

17. Girshick R, Donahue J, Darrell T, Malik J (2014) Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp.580 587)

18. He K, Zhang X, Ren S, Sun J (2016) Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition(pp.770 778)