o oi@shddlmZddlmZddlZddlmZmZddlm Z ddl m Z ddl m Z Gd d d eZdS) ) annotations)OptionalN)ModuleTensor)Unet)heatmap_to_keypoints) DISKFeaturescsReZdZdZd%d&fd d Zd'ddZ    d(d)ddZed*d+d#d$ZZ S),DISKa!Module which detects and described local features in an image using the DISK method. See :cite:`tyszkiewicz2020disk` for details. .. image:: _static/img/disk_outdoor_depth.jpg Args: desc_dim: The dimension of the descriptor. unet: The U-Net to use. If None, a default U-Net is used. Kornia doesn't provide the training code for DISK so this is only useful when using a custom checkpoint trained using the code released with the paper. The unet should take as input a tensor of shape :math:`(B, C, H, W)` and output a tensor of shape :math:`(B, \mathrm{desc\_dim} + 1, H, W)`. Example: >>> disk = DISK.from_pretrained('depth') >>> images = torch.rand(1, 3, 256, 256) >>> features = disk(images) Ndesc_dimintunet None | ModulereturnNonec sBt||_|durtddgdddd|dgd}||_dS)N) @rrrr) in_featuressizedownup)super__init__r rr)selfr r __class__L/home/ubuntu/.local/lib/python3.10/site-packages/kornia/feature/disk/disk.pyr4s   z DISK.__init__imagesrtuple[Tensor, Tensor]cCsp||}|jd|jdkrtd|jdd|jd|ddd|jf}|dd|jdf}||fS)aReturn the heatmap and the dense descriptors. .. image:: _static/img/DISK.png Args: images: The image to detect features in. Shape :math:`(B, 3, H, W)`. Returns: A tuple of dense detection scores and descriptors. Shapes are :math:`(B, 1, H, W)` and :math:`(B, D, H, W)`, where :math:`D` is the descriptor dimension. rzU-Net output has z& channels, but expected self.desc_dim=z + 1.N)rshaper ValueError)rr" unet_output descriptorsheatmapsr r r!heatmap_and_dense_descriptors=s z"DISK.heatmap_and_dense_descriptorsrFn Optional[int] window_sizescore_thresholdfloatpad_if_not_divisibleboollist[DISKFeatures]cCs|jd}|r:|jdd\}}|ddkrd|dnd} |ddkr*d|dnd} tjjj|d| d| fdd}||\} } |rY| dd|d|f} | dd|d|f} t| |||d} g}t|D]}|| | | |qg|S) aDetect features in an image, returning keypoint locations, descriptors and detection scores. Args: images: The image to detect features in. Shape :math:`(B, 3, H, W)`. n: The maximum number of keypoints to detect. If None, all keypoints are returned. window_size: The size of the non-maxima suppression window used to filter detections. score_threshold: The minimum score a detection must have to be returned. See :py:class:`DISKFeatures` for details. pad_if_not_divisible: if True, the non-16 divisible input is zero-padded to the closest 16-multiply Returns: A list of length :math:`B` containing the detected features. rNrr*)value.)r+r-r.) r$torchnn functionalpadr)rrangeappendmerge_with_descriptors)rr"r+r-r.r0Bhwpd_hpd_wr(r' keypointsfeaturesir r r!forwardWs  z DISK.forwarddepth checkpointstrdeviceOptional[torch.device]cCslddd}||vrtd||durtd}tjj|||d}||}||d||S) a~Load a pretrained model. Depth model was trained using depth map supervision and is slightly more precise but biased to detect keypoints only where SfM depth is available. Epipolar model was trained using epipolar geometry supervision and is less precise but detects keypoints everywhere where they are matchable. The difference is especially pronounced on thin structures and on edges of objects. Args: checkpoint: The checkpoint to load. One of 'depth' or 'epipolar'. device: The device to load the model to. Returns: The pretrained model. zGhttps://raw.githubusercontent.com/cvlab-epfl/disk/master/depth-save.pthzJhttps://raw.githubusercontent.com/cvlab-epfl/disk/master/epipolar-save.pth)rEepipolarzUnknown pretrained model: Ncpu) map_location extractor)r%r5rHhubload_state_dict_from_urltoload_state_dicteval)clsrFrHurlspretrained_dictmodelr r r!from_pretraineds  zDISK.from_pretrained)r N)r r rrrr)r"rrr#)Nrr*F) r"rr+r,r-r r.r/r0r1rr2)rEN)rFrGrHrIrr ) __name__ __module__ __qualname____doc__rr)rD classmethodrW __classcell__r r rr!r s  *r ) __future__rtypingrr5 kornia.corerr_unetsrdetectorrstructsr r r r r r!s