3j$SSKJr SSKrSSKrSSKJr SSKJrJrJrJ r SSK J r J r SSK JrJr S SjrS SjrS S S jjr"S S \R&5rg)) annotationsN)nn) KORNIA_CHECKKORNIA_CHECK_IS_TENSORKORNIA_CHECK_SHAPEKORNIA_CHECK_TYPE)filter2dfilter3d)create_meshgridcreate_meshgrid3dc.URnURnUS-n[R"[ UR SUR S5U- 5n[ XSX45nXu- n[R"USUS5n[R"U*U- 5RS5n [R"U5n UR5n [R"U 5n [U5Hn [XSS9nU*[R "U5-n[R""US S S S 9nUS:nUR%5(d U $X-nU UU-UR'U RS 9--n [R("XU 5n M U $) NF)r.r)r.r replicate border_typenanposinfneginfdtype)devicermathceilmaxshaper torchhypotexp unsqueeze zeros_likeclone ones_likeranger log nan_to_numanytowhereimage kernel_sizehrrk_halfn_itersgriddistkerneloutboundary signal_onesicdtmaskoffsets [/home/wildlama/miniconda3/lib/python3.13/site-packages/kornia/contrib/distance_transform.py_distance_transform_2d_implr>sa \\F KKE A FiiEKKNEKKN;fDEG ;UF JD =D ;;tIY 8D YYuqy ! + +A .F   5 !C{{}H//(+K 7^x[Ab599S>!sCDQwxxzz  J jVc\TWW399W%===;;t(; Jc URnURnUS-n[R"[ UR SS5U- 5n[ XUSX45nXu- n[R"USSSS9n[R"U*U- 5RS5n [R"U5n UR5n [R"U 5n [U5Hn [XSS9nU*[R "U5-n[R""USSSS 9nUS:nUR%5(d U $X-nU UU-UR'U RS 9--n [R("XU 5n M U $) NrFr)pdimrrrrr)rrrrrrr r normr"r#r$r%r&r'r r(r)r*r+r,r-s r=_distance_transform_3d_implrE>sY \\F KKE A FiiEKKO,v56G [{E6 YD =D ::d1g +D YYuqy ! + +A .F   5 !C{{}H//(+K 7^x[Ab599S>!sCDQwxxzz  J jVc\TWW399W%===;;t(; Jr?c[U5 [UR5S5 [URS;SUR35 URS:Xa[ U/SQ5 O[ U/SQ5 [ U[S5 [US-S :g=(a US :S 5 [US :S U35 URS:Xa [XU5$[XU5$) asApproximates the Euclidean distance transform of images/volumes using cascaded convolution operations. The value at each pixel/voxel represents the distance to the nearest non-zero element. It uses the method described in :cite:`pham2021dtlayer`. The transformation is applied independently across the channel dimension. Args: image: Image or volume with shape :math:`(B,C,H,W)` or :math:`(B,C,D,H,W)`. kernel_size: size of the convolution kernel. Must be an odd number. h: value that influence the approximation of the min function. Returns: tensor with the same shape as input. Example: >>> # 2D example: >>> tensor = torch.zeros(1, 1, 5, 5) >>> tensor[:,:, 1, 2] = 1 >>> dt = distance_transform(tensor) >>> # 3D example: >>> volume = torch.zeros(1, 1, 5, 5, 5) >>> volume[:, :, 2, 2, 2] = 1 >>> dt = distance_transform(volume) z%image must be a floating point tensor)z:Invalid image shape, we expect BxCxHxW or BxCxDxHxW. Got: rG)BCHW)rIrJDrKrLzkernel_size must be an intrrrz'kernel_size must be an odd integer >= 3z h must be a positive float, got ) rris_floating_pointndimrrrintr>rE)r.r/r0s r=distance_transformrQ]s65!((*,STv%)cdidodocp'qr zzQ5"675";<k3(DEqA%:+*:? zzQ*5qAA &u1 ==r?c>^\rSrSrSrSSU4SjjjrSSjrSrU=r$) DistanceTransformzModule that approximates the Euclidean distance transform of images/volumes using convolutions. Args: kernel_size: size of the convolution kernel. h: value that influence the approximation of the min function. c:>[TU]5 XlX lg)N)super__init__r/r0)selfr/r0 __class__s r=rWDistanceTransform.__init__s &r?cURSS:a$URSSnUR"SS/UQ76nOUn[X0RUR5R U5$)aCompute the distance transform while preserving the original tensor layout. The underlying functional implementation expects a single-channel tensor. For multi-channel inputs, channels are temporarily folded into the batch dimension, processed independently, and then reshaped back. Args: image: Input tensor with shape :math:`(B, C, H, W)` for 2D data or :math:`(B, C, D, H, W)` for 3D data, where: - ``B`` is batch size, - ``C`` is the number of channels, - ``D`` is depth (only for 3D volumes), - ``H`` is height, - ``W`` is width. Returns: A distance-transform tensor with the same shape as ``image``. Each channel is processed independently. rrNrA)rreshaperQr/r0view_as)rXr. spatial_dimsimage_ins r=forwardDistanceTransform.forwards_* ;;q>A  ;;qr?L}}R:\:HH!(,<,rErQModulerSr?r=rtsB$# ii-C@>->`+U +Ur?