ZipDepth B

Click a block to read its description, or select it with Tab and Enter.

ZipDepth BRelative inverse depth,input3 × 384 × 384,native unfused eval,global_mode=balanced. Shapes exclude batch.LibreYOLOZipDepth BRelative inverse depth,input3 × 384 × 384,native unfused eval,global_mode=balanced. Shapes exclude batch.EncoderRGB + ImageNet normalization3 × 384 × 384ConvBN 3×3,s23 to24;half skip24×192²ConvBN3×3,s224 to48;48×96²QARepBlock,n=248 to48;S1:48×96²QARepBlock stride248 to96;48²QARepBlock,n=296 to96;48²MinimalMultiScale thenStripPooling96×48²;S2QARepBlock stride296 to192;24²QARepBlock,n=6192 to192;24²ChannelAttention thenGlobalContext192×24²;S3QARepBlock stride2192 to384;12²QARepBlock,n=2384×12²LightweightSPPF384×12²;S4Cross-scale exchange and FPNMinimalCrossScale(S3,S4)S3:192×24², S4:384×12²ConvBN1×1 on exchangedS4384 to288;12²UltraLightFusion3high192/low288 to192;24²UltraLightFusion2high96/low192 to144;48²UltraLightFusion1high48/low144 to96;96²UltraLightFusion halfhigh24/low96 to32;192²High-resolution inputs come fromS3,S2,S1 andhalf-stem skip.Half-resolution depth and final2× reconstructionHalf feature32×192×192Conv2d3×332 to1;s1,p1;halfdepth192²Convex upsampleb:9-neighbor convex. bnpu:learnedinterpolation alpha.ReLUNonnegative relative inverse depthOutput1×384×384Both trained heads also read the32-channel half feature.b andbnpu have different head parameters; no weight-free rewrite.Public loaded models fuse QARep branches; this view is unfused.QARepBlockStage widths48,96,192,384. Transitions48to96,96to192,192to384.Conv2d3×3Ci toCo;declaredstride;biasFalseBatchNorm2dCo channelsConv2d1×1Ci toCo;declaredstride;biasFalseBatchNorm2dCo channels++Input identityOnly Ci=Co,s1ReLUOne fusedConv3×3 can replacebranches afterloadIdentity branch has no BatchNorm; downsample blocks omit it.ChannelAttentionInput192×24²Mean overH,W192×1×1Conv1×1192to24;biasFalseReLUConv1×124to192;biasFalseSigmoid×StripPoolingAttentionMean overwidth96×48×1Mean overheight96×1×48+DepthwiseConv1×196channels/groups;biasFalseBatchNorm2d96channelsSigmoid×Input XMinimalMultiScaleDepthwiseConv3×396groups;dilation1,padding1DepthwiseConv3×396groups;dilation2,padding2+BatchNorm2d96 channels+Input X96×48²Residual addition has no finalactivation.GlobalContextBlockConv1×1;flatten192to1;576spatial logitsSoftmax over576positionsContext weightsBatch matmul X × weights192×576 times576×1Conv1×1192to48BatchNorm2d + ReLU48channelsConv1×148to192+Input XLightweightSPPFConvBN1×1384to96;12²MaxPool5×5,stride1,padding296×12²MaxPool5×5,stride1,padding296×12²MaxPool5×5,stride1,padding296×12²xp1p2p3Concat x,p1,p2,p3384×12²ConvBN1×1384to384Bidirectional MinimalCrossScaleGroupedConv1×1384to192;groups4;biasFalseNearest resize12² to24²Match targetgridMultiply0.3+OriginalS3GroupedConv1×1192to384;groups4;biasFalseAvgPool2×2,stride2Match targetgridMultiply0.3+OriginalS4Both exchange paths read originalS3/S4, before either result is updated.UltraLightFusion and ConvBNHigh-resolution sourceChannels192,96,48,24 across four fusionsBilinear resize low-resolution sourceChannels288,192,144,96;align_cornersFalseGroupedConv1×1,groups4Both project to192,144,96,32 respectivelyGroupedConv1×1,groups4Both project to192,144,96,32 respectively+BatchNorm2d thenReLU192,144,96,32channels respectivelyConvBN elsewhere means Conv2d (biasFalse),BatchNorm2d,ReLU.b:9-neighbor convex headConv3×3 +BN +ReLU32to8;192²;biasFalseConv1×18to36 =9neighbors×4subpixelsReshape andsoftmax over9neighbors9×4×192×192;temperature1Replicate-pad halfdepth;Unfold3×39×1×192×192 neighborhood valuesMultiply masks × neighbors,sum94×192×192PixelShuffle21×384×384DepthBoth heads finish withReLU. They are separately trained checkpoint variants.bnpu:unfold-free trained headConv1×1 +BN +ReLU32to16;192²;biasFalseDepthwiseConv5×5 +BN +ReLU16groups;padding2;192²Conv1×1;bilinear×2;sigmoid16to1;alpha1×384×384Nearest×2 andbilinear×2 halfdepthTwo1×384×384 candidate mapsalpha×nearest + (1-alpha)×bilinearLearned convex interpolationDepthBoth heads finish withReLU. They are separately trained checkpoint variants.Balanced mode includes strip pooling andGC context; it does not execute the optional full-mode global-token attention.Source: libreyolo/models/zipdepth/nn.py and model.py. Revision a4d0ecc9e17f.libreyolo.com