Octree Generating Networks_ Efficient Convolutional Architectures for High-Resolution 3D Outputs

《Octree Generating Networks_ Efficient Convolutional Architectures for High-Resolution 3D Outputs》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、OctreeGeneratingNetworks:EfficientConvolutionalArchitecturesforHigh-resolution3DOutputsMaximTatarchenko1AlexeyDosovitskiy1,2ThomasBrox1tatarchm@cs.uni-freiburg.deadosovitskiy@gmail.combrox@cs.uni-freiburg.de1UniversityofFreiburg2IntelLabsAbstractOctreeOctreeOctreedenselevel1level2level3Weprese

2、ntadeepconvolutionaldecoderarchitecturethatcangeneratevolumetric3Doutputsinacompute-andmemory-efficientmannerbyusinganoctreerepresentation.Thenetworklearnstopredictboththestructureoftheoc-tree,andtheoccupancyvaluesofindividualcells.Thismakesitaparticularlyvaluabletechniqueforgenerating32364312

3、833Dshapes.Incontrasttostandarddecodersactingonreg-Figure1.TheproposedOGNrepresentsitsvolumetricoutputasanularvoxelgrids,thearchitecturedoesnothavecubiccom-octree.Initiallyestimatedroughlow-resolutionstructureisgradu-plexity.Thisallowsrepresentingmuchhigherresolutionallyrefinedtoadesiredhighre

4、solution.Ateachlevelonlyasparseoutputswithalimitedmemorybudget.Wedemonstratethissetofspatiallocationsispredicted.Thisrepresentationissignifi-inseveralapplicationdomains,including3Dconvolutionalcantlymoreefficientthanadensevoxelgridandallowsgeneratingautoencoders,generationofobjectsandwholescene

5、sfromvolumesaslargeas5123voxelsonamodernGPUinasinglefor-high-levelrepresentations,andshapefromasingleimage.wardpass.largecellofanoctree,resultinginsavingsincomputation1.Introductionandmemorycomparedtoafineregulargrid.Atthesametime,finedetailsarenotlostandcanstillberepresentedbyUp-convolutional1

6、decoderarchitectureshavebecomesmallcellsoftheoctree.astandardtoolfortasksrequiringimagegeneration[9,Wepresentanoctreegeneratingnetwork(OGN)-acon-27,20]orper-pixelprediction[25,8].Theyconsistofvolutionaldecoderoperatingonoctrees.Thecoarsestruc-aseriesofconvolutionalandup-convolutional(upsam-tu

7、reofthenetworkisillustratedinFigure1.Similartoapling+convolution)layersoperatingonregulargrids,withusualup-convolutionaldecoder,therepresentationisgrad-resolutiongraduallyincreasingtowardstheoutputoftheuallyconvolvedwithlearnedfiltersandup-sam