a d [@shddlZddlmZmZddlZddlmZgdZejj ej e dddZ ejj ej e dd d Z ej eej ej fdd d Zej eej ej ej fdd dZej eeej dddZej eeeefej dddZej ej ej ej ej dddZej ej ej ej ej ej ej dddZd"ej ej eee ej dddZej ej eej dd d!ZdS)#N)OptionalTuple)transform_points) validate_bboxvalidate_bbox3dinfer_bbox_shapeinfer_bbox_shape3d bbox_to_maskbbox_to_mask3dbbox_generatorbbox_generator3dtransform_bboxnms)boxesreturnc Cst|jdvr*|jddtddgks>> boxes = torch.tensor([[ ... [1., 1.], ... [2., 1.], ... [2., 2.], ... [1., 2.], ... ], [ ... [1., 1.], ... [3., 1.], ... [3., 2.], ... [1., 2.], ... ]]) # 2x4x2 >>> infer_bbox_shape(boxes) (tensor([2., 2.]), tensor([2., 3.])) Nrrr)r)rwidthheightr"r"r#rWs((rc CsNt|t|dtjgd|jtjddddddf}t|dtjgd|jtjddddddf}||ddddf}t|dtjgd|jtjddddddf}t|dtjgd|jtjddddddf}||ddddf}|ddd dd f|dddd d fddddf}|||fS) aKAuto-infer the output sizes for the given 3D bounding boxes. Args: boxes: a tensor containing the coordinates of the bounding boxes to be extracted. The tensor must have the shape of Bx8x3, where each box is defined in the following ``clockwise`` order: front-top-left, front-top-right, front-bottom-right, front-bottom-left, back-top-left, back-top-right, back-bottom-right, back-bottom-left. The coordinates must be in the x, y, z order. Returns: - Bounding box depths, shape of :math:`(B,)`. - Bounding box heights, shape of :math:`(B,)`. - Bounding box widths, shape of :math:`(B,)`. Example: >>> boxes = torch.tensor([[[ 0, 1, 2], ... [10, 1, 2], ... [10, 21, 2], ... [ 0, 21, 2], ... [ 0, 1, 32], ... [10, 1, 32], ... [10, 21, 32], ... [ 0, 21, 32]], ... [[ 3, 4, 5], ... [43, 4, 5], ... [43, 54, 5], ... [ 3, 54, 5], ... [ 3, 4, 65], ... [43, 4, 65], ... [43, 54, 65], ... [ 3, 54, 65]]]) # 2x8x3 >>> infer_bbox_shape3d(boxes) (tensor([31, 61]), tensor([21, 51]), tensor([11, 41])) rr%r(Nrr+r-r.rr)rrr/r0r)r1r2r"r"r#r xs"4444<r )rr6r7rcCst|tjt||d|dftj|jd}|d}t||D]4\}}d||d|dd|d|ddf<qD|d d dd dd fS) aConvert 2D bounding boxes to masks. Covered area is 1. and the remaining is 0. Args: boxes: a tensor containing the coordinates of the bounding boxes to be extracted. The tensor must have the shape of Bx4x2, where each box is defined in the following ``clockwise`` order: top-left, top-right, bottom-right and bottom-left. The coordinates must be in the x, y order. width: width of the masked image. height: height of the masked image. Returns: the output mask tensor. Note: It is currently non-differentiable. Examples: >>> boxes = torch.tensor([[ ... [1., 1.], ... [3., 1.], ... [3., 2.], ... [1., 2.], ... ]]) # 1x4x2 >>> bbox_to_mask(boxes, 5, 5) tensor([[[0., 0., 0., 0., 0.], [0., 1., 1., 1., 0.], [0., 1., 1., 1., 0.], [0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]]]) r)r*r)rg?rrrrrrrrNr)rrzerosrfloatr)r1zip)rr6r7maskZbox_iZmskbxr"r"r#r s & .r )rsizerc Cst|tt|g|R}g}t||D]T\}}|dtj|d|dd|jtj dtj d|j|j d}|dtj|d|dd|jtj dtj d|j|j d}|dtj|d |d d|jtj dtj d|j|j d}|j dd }t |}||dkjdd d jdd d }||dkjdd d jdd d }||dkjdd d jdd d }||q,tj|dd S)aHConvert 3D bounding boxes to masks. Covered area is 1. and the remaining is 0. Args: boxes: a tensor containing the coordinates of the bounding boxes to be extracted. The tensor must have the shape of Bx8x3, where each box is defined in the following ``clockwise`` order: front-top-left, front-top-right, front-bottom-right, front-bottom-left, back-top-left, back-top-right, back-bottom-right, back-bottom-left. The coordinates must be in the x, y, z order. size: depth, height and width of the masked image. Returns: the output mask tensor. Examples: >>> boxes = torch.tensor([[ ... [1., 1., 1.], ... [2., 1., 1.], ... [2., 2., 1.], ... [1., 2., 1.], ... [1., 1., 2.], ... [2., 1., 2.], ... [2., 2., 2.], ... [1., 2., 2.], ... ]]) # 1x8x3 >>> bbox_to_mask3d(boxes, (4, 5, 5)) tensor([[[[[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]], [[0., 0., 0., 0., 0.], [0., 1., 1., 0., 0.], [0., 1., 1., 0., 0.], [0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]], [[0., 0., 0., 0., 0.], [0., 1., 1., 0., 0.], [0., 1., 1., 0., 0.], [0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]], [[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]]]]]) r)rr)rrrr()rrr9rr:r;dimT)rCZkeepdimr)rrr<rr>Z index_fillZarangeitemr)r1r0r* unsqueezeZ ones_likeallappendstackr=)rrAr?Zmask_outmZboxZm_outr"r"r#r s41(((      r )x_starty_startr6r7rc Cs|j|jkr|dvs.td|d|d|j|jkrF|dvs\td|d|d|j|jkr|jkr|jksntd|jd|jd|jd |jd |j|jkr|jkr|jksntd |jd|jd|jd |jd tjd d gd d gd d gd d ggg|j|jd |d krBdnt|dd}|ddddd f| dd7<|dddddf| dd7<|dddd f|d7<|dddd f|d7<|ddddf|d7<|ddddf|d7<|S)aNGenerate 2D bounding boxes according to the provided start coords, width and height. Args: x_start: a tensor containing the x coordinates of the bounding boxes to be extracted. Shape must be a scalar tensor or :math:`(B,)`. y_start: a tensor containing the y coordinates of the bounding boxes to be extracted. Shape must be a scalar tensor or :math:`(B,)`. width: widths of the masked image. Shape must be a scalar tensor or :math:`(B,)`. height: heights of the masked image. Shape must be a scalar tensor or :math:`(B,)`. Returns: the bounding box tensor. Examples: >>> x_start = torch.tensor([0, 1]) >>> y_start = torch.tensor([1, 0]) >>> width = torch.tensor([5, 3]) >>> height = torch.tensor([7, 4]) >>> bbox_generator(x_start, y_start, width, height) tensor([[[0, 1], [4, 1], [4, 7], [0, 7]], [[1, 0], [3, 0], [3, 3], [1, 3]]]) r8z6`x_start` and `y_start` must be a scalar or (B,). Got , rz3`width` and `height` must be a scalar or (B,). Got 5All tensors must be in the same dtype. Got `x_start`( ), `y_start`( ), `width`( ), `height`().6All tensors must be in the same device. Got `x_start`(rr(rNrrr) rrCrr*r)rr0repeatrr)rJrKr6r7bboxr"r"r#r "sL (*.&&r )rJrKz_startr6r7depthrcCs|j|jkr|jkr*nn |dvsFtd|d|d|d|j|jkr`|jkrpnn |dvstd|d|d|d|j|jkr|jkr|jkr|jkr|jksntd|jd|jd|jd |jd |jd |jd |j|jkrP|jkrP|jkrP|jkrP|jksntd |jd|jd|jd |jd |jd |jd tjgdgdgdgdgg|j|jdt|dd}|dddddf| dd7<|dddddf| dd7<|dddddf| dd7<|ddddf|7<|ddddf|7<|ddddf|7<|ddddf|7<| }|dddddf|j dddd7<tj ||gdd}|S)aGenerate 3D bounding boxes according to the provided start coords, width, height and depth. Args: x_start: a tensor containing the x coordinates of the bounding boxes to be extracted. Shape must be a scalar tensor or :math:`(B,)`. y_start: a tensor containing the y coordinates of the bounding boxes to be extracted. Shape must be a scalar tensor or :math:`(B,)`. z_start: a tensor containing the z coordinates of the bounding boxes to be extracted. Shape must be a scalar tensor or :math:`(B,)`. width: widths of the masked image. Shape must be a scalar tensor or :math:`(B,)`. height: heights of the masked image. Shape must be a scalar tensor or :math:`(B,)`. depth: depths of the masked image. Shape must be a scalar tensor or :math:`(B,)`. Returns: the 3d bounding box tensor :math:`(B, 8, 3)`. Examples: >>> x_start = torch.tensor([0, 3]) >>> y_start = torch.tensor([1, 4]) >>> z_start = torch.tensor([2, 5]) >>> width = torch.tensor([10, 40]) >>> height = torch.tensor([20, 50]) >>> depth = torch.tensor([30, 60]) >>> bbox_generator3d(x_start, y_start, z_start, width, height, depth) tensor([[[ 0, 1, 2], [10, 1, 2], [10, 21, 2], [ 0, 21, 2], [ 0, 1, 32], [10, 1, 32], [10, 21, 32], [ 0, 21, 32]], [[ 3, 4, 5], [43, 4, 5], [43, 54, 5], [ 3, 54, 5], [ 3, 4, 65], [43, 4, 65], [43, 54, 65], [ 3, 54, 65]]]) r8zA`x_start`, `y_start` and `z_start` must be a scalar or (B,). Got rLrz<`width`, `height` and `depth` must be a scalar or (B,). Got rMrNz ), `z_start`(rOrPz) and `depth`(rQrR)rrrr(rNrrrrrBr) rrCrr*r)rr0rSrrclonerEcat)rJrKrUr6r7rVrTZ bbox_backr"r"r#r `sl2**BJ$ &&&.r xyxy) trans_matrmoderestore_coordinatesrcCst|tstdt||dvr2td||dur`|jddtddgks`t d|d kr|d |d |d <|d |d |d <t || |jddd}| |}|dus|rt|jddtddgkst| }tj|dddgfddd|d <tj|dddgfddd|d <tj|dddgfddd|d <tj|dddgfddd|d <|}|d kr|d |d |d <|d |d |d <|S)aqApply a transformation matrix to a box or batch of boxes. Args: trans_mat: The transformation matrix to be applied with a shape of :math:`(3, 3)` or batched as :math:`(B, 3, 3)`. boxes: The boxes to be transformed with a common shape of :math:`(N, 4)` or batched as :math:`(B, N, 4)`, the polygon shape of :math:`(B, N, 4, 2)` is also supported. mode: The format in which the boxes are provided. If set to 'xyxy' the boxes are assumed to be in the format ``xmin, ymin, xmax, ymax``. If set to 'xywh' the boxes are assumed to be in the format ``xmin, ymin, width, height`` restore_coordinates: In case the boxes are flipped, adding a post processing step to restore the coordinates to a valid bounding box. Returns: The set of transformed points in the specified mode zMode must be a string. Got )rYxywhz(Mode must be one of 'xyxy', 'xywh'. Got NrrraPPrevious behaviour produces incorrect box coordinates if a flip transformation performed on boxes.The previous wrong behaviour has been corrected and will be removed in the future versions.If you wish to keep the previous behaviour, please set `restore_coordinates=False`.Otherwise, set `restore_coordinates=True` as an acknowledgement.r]).r).r).r).rrr.rBrr) isinstancer! TypeErrortyper rrrwarningswarnrrZview_asrWminmax)rZrr[r\Ztransformed_boxesZrestored_boxesr"r"r#rs0 $ ,"""" r)rscores iou_thresholdrcCst|jdkr.|jddkr.td|jdt|jdkrNtd|jd|jd|jdkrztd |j|jfd|d\}}}}||||}|jd d \}} g} | jddkr| d} | | t|| || dd } t|| || dd } t|| || dd }t|| || dd }tj || d d}tj || d d}||}||| || dd |}t ||kd}| |d} qt| dkrt | St | S)aaPerform non-maxima suppression (NMS) on a given tensor of bounding boxes according to the intersection-over- union (IoU). Args: boxes: tensor containing the encoded bounding boxes with the shape :math:`(N, (x_1, y_1, x_2, y_2))`. scores: tensor containing the scores associated to each bounding box with shape :math:`(N,)`. iou_threshold: the throshold to discard the overlapping boxes. Return: A tensor mask with the indices to keep from the input set of boxes and scores. Example: >>> boxes = torch.tensor([ ... [10., 10., 20., 20.], ... [15., 5., 15., 25.], ... [100., 100., 200., 200.], ... [100., 100., 200., 200.]]) >>> scores = torch.tensor([0.9, 0.8, 0.7, 0.9]) >>> nms(boxes, scores, iou_threshold=0.8) tensor([0, 3, 1]) rrrzboxes expected as Nx4. Got: rrzscores expected as N. Got: rz+boxes and scores mus have same shape. Got: T)Z descendingNg)rc) rrr ZunbindsortrGrrdrcclampwhererHr0)rrerfx1y1Zx2y2Zareas_orderZkeepiZxx1Zyy1Zxx2Zyy2whinterZovrZindsr"r"r#rs4   r)rYN)ratypingrrrZlinalgr__all__ZjitignoreZTensorboolrrrr intr r r r r!rr=rr"r"r"r#s: "!"0* Q ? [ ;