neRCNN：从单张图像检测重建平面区域.pdfVIP

Figure2.Ourframeworkconsistsofthreebuildingblocks:1)aplanedetectionnetworkbasedonMaskR-CNN[14],2)asegmentation

refinementnetworkthatjointlyoptimizesextractedsegmentationmasks,and3)awarpinglossmodulethatenforcestheconsistencyof

reconstructionswithanearbyviewduringtraining.

egories(e.g.,person,car,train,bicycleandmore).Ouroneseparatefully-connectedlayertoregressthe3Dresid-

problemhasonlytwocategories”planar”or”non-planar”,ualvectorforeachanchornormal(i.e.,21=3×7out-

defi