Source code for kornia.feature.adalam.adalam

# Integrated from original AdaLAM repo
# https://github.com/cavalli1234/AdaLAM
# Copyright (c) 2020, Luca Cavalli

from typing import Dict, Optional, Tuple

import torch
from torch import Tensor

from kornia.feature.laf import get_laf_center, get_laf_orientation, get_laf_scale
from kornia.testing import KORNIA_CHECK_LAF, KORNIA_CHECK_SHAPE
from kornia.utils.helpers import get_cuda_device_if_available

from .core import adalam_core
from .utils import dist_matrix


def get_adalam_default_config():
    DEFAULT_CONFIG = {
        'area_ratio': 100,  # Ratio between seed circle area and image area. Higher values produce more seeds with smaller neighborhoods.    # noqa: E501
        'search_expansion': 4,  # Expansion factor of the seed circle radius for the purpose of collecting neighborhoods. Increases neighborhood radius without changing seed distribution    # noqa: E501
        'ransac_iters': 128,  # Fixed number of inner GPU-RANSAC iterations
        'min_inliers': 6,  # Minimum number of inliers required to accept inliers coming from a neighborhood    # noqa: E501
        'min_confidence': 200,  # Threshold used by the confidence-based GPU-RANSAC
        'orientation_difference_threshold': 30,  # Maximum difference in orientations for a point to be accepted in a neighborhood. Set to None to disable the use of keypoint orientations.   # noqa: E501
        'scale_rate_threshold': 1.5,  # Maximum difference (ratio) in scales for a point to be accepted in a neighborhood. Set to None to disable the use of keypoint scales.   # noqa: E501
        'detected_scale_rate_threshold': 5,  # Prior on maximum possible scale change detectable in image couples. Affinities with higher scale changes are regarded as outliers.   # noqa: E501
        'refit': True,  # Whether to perform refitting at the end of the RANSACs. Generally improves accuracy at the cost of runtime.   # noqa: E501
        'force_seed_mnn': True,  # Whether to consider only MNN for the purpose of selecting seeds. Generally improves accuracy at the cost of runtime.    # noqa: E501
        # You can provide a MNN mask in input to skip MNN computation and still get the improvement.
        'device': get_cuda_device_if_available(),  # Device to be used for running AdaLAM. Use GPU if available.   # noqa: E501
    }
    return DEFAULT_CONFIG


[docs]def match_adalam( desc1: Tensor, desc2: Tensor, lafs1: Tensor, lafs2: Tensor, config: Optional[Dict] = None, hw1: Optional[Tensor] = None, hw2: Optional[Tensor] = None, dm: Optional[Tensor] = None, ) -> Tuple[Tensor, Tensor]: """Function, which performs descriptor matching, followed by AdaLAM filtering (see :cite:`AdaLAM2020` for more details) If the distance matrix dm is not provided, :py:func:`torch.cdist` is used. Args: desc1: Batch of descriptors of a shape :math:`(B1, D)`. desc2: Batch of descriptors of a shape :math:`(B2, D)`. lafs1: LAFs of a shape :math:`(1, B1, 2, 3)`. lafs2: LAFs of a shape :math:`(1, B1, 2, 3)`. config: dict with AdaLAM config dm: Tensor containing the distances from each descriptor in desc1 to each descriptor in desc2, shape of :math:`(B1, B2)`. Return: - Descriptor distance of matching descriptors, shape of :math:`(B3, 1)`. - Long tensor indexes of matching descriptors in desc1 and desc2. Shape: :math:`(B3, 2)`, where 0 <= B3 <= B1. """ KORNIA_CHECK_SHAPE(desc1, ["B", "DIM"]) KORNIA_CHECK_SHAPE(desc2, ["B", "DIM"]) KORNIA_CHECK_LAF(lafs1) KORNIA_CHECK_LAF(lafs2) if config is None: config_ = get_adalam_default_config() config_['device'] = desc1.device else: config_ = config adalam_object = AdalamFilter(config_) idxs = adalam_object.match_and_filter( get_laf_center(lafs1).reshape(-1, 2), get_laf_center(lafs2).reshape(-1, 2), desc1, desc2, hw1, hw2, get_laf_orientation(lafs1).reshape(-1), get_laf_orientation(lafs2).reshape(-1), get_laf_scale(lafs1).reshape(-1), get_laf_scale(lafs2).reshape(-1), ) quality = torch.ones(len(idxs), 1, dtype=desc1.dtype, device=desc1.device) return quality, idxs
class AdalamFilter: DEFAULT_CONFIG = get_adalam_default_config() def __init__(self, custom_config: dict = None): """This class acts as a wrapper to the method AdaLAM for outlier filtering. init args: custom_config: dictionary overriding the default configuration. Missing parameters are kept as default. See documentation of DEFAULT_CONFIG for specific explanations on the accepted parameters. """ self.config = AdalamFilter.DEFAULT_CONFIG.copy() if custom_config is not None: for key, val in custom_config.items(): if key not in self.config.keys(): print( "WARNING: custom configuration contains a key which is not recognized ({key}). " "Known configurations are {list(self.config.keys())}." ) continue self.config[key] = val def filter_matches( self, k1: torch.Tensor, k2: torch.Tensor, putative_matches: torch.Tensor, scores: torch.Tensor, mnn: torch.Tensor = None, im1shape: tuple = None, im2shape: tuple = None, o1: torch.Tensor = None, o2: torch.Tensor = None, s1: torch.Tensor = None, s2: torch.Tensor = None, ): """Call the core functionality of AdaLAM, i.e. just outlier filtering. No sanity check is performed on the inputs. Inputs: k1: keypoint locations in the source image, in pixel coordinates. Expected a float32 tensor with shape (num_keypoints_in_source_image, 2). k2: keypoint locations in the destination image, in pixel coordinates. Expected a float32 tensor with shape (num_keypoints_in_destination_image, 2). putative_matches: Initial set of putative matches to be filtered. The current implementation assumes that these are unfiltered nearest neighbor matches, so it requires this to be a list of indices a_i such that the source keypoint i is associated to the destination keypoint a_i. For now to use AdaLAM on different inputs a workaround on the input format is required. Expected a long tensor with shape (num_keypoints_in_source_image,). scores: Confidence scores on the putative_matches. Usually holds Lowe's ratio scores. mnn: A mask indicating which putative matches are also mutual nearest neighbors. See documentation on 'force_seed_mnn' in the DEFAULT_CONFIG. If None, it disables the mutual nearest neighbor filtering on seed point selection. Expected a bool tensor with shape (num_keypoints_in_source_image,) im1shape: Shape of the source image. If None, it is inferred from keypoints max and min, at the cost of wasted runtime. So please provide it. Expected a tuple with (width, height) or (height, width) of source image im2shape: Shape of the destination image. If None, it is inferred from keypoints max and min, at the cost of wasted runtime. So please provide it. Expected a tuple with (width, height) or (height, width) of destination image o1/o2: keypoint orientations in degrees. They can be None if 'orientation_difference_threshold' in config is set to None. See documentation on 'orientation_difference_threshold' in the DEFAULT_CONFIG. Expected a float32 tensor with shape (num_keypoints_in_source/destination_image,) s1/s2: keypoint scales. They can be None if 'scale_rate_threshold' in config is set to None. See documentation on 'scale_rate_threshold' in the DEFAULT_CONFIG. Expected a float32 tensor with shape (num_keypoints_in_source/destination_image,) Returns: Filtered putative matches. A long tensor with shape (num_filtered_matches, 2) with indices of corresponding keypoints in k1 and k2. """ # noqa: E501 with torch.no_grad(): return adalam_core( k1, k2, fnn12=putative_matches, scores1=scores, mnn=mnn, im1shape=im1shape, im2shape=im2shape, o1=o1, o2=o2, s1=s1, s2=s2, config=self.config, ) def match_and_filter(self, k1, k2, d1, d2, im1shape=None, im2shape=None, o1=None, o2=None, s1=None, s2=None): """Standard matching and filtering with AdaLAM. This function: - performs some elementary sanity check on the inputs; - wraps input arrays into torch tensors and loads to GPU if necessary; - extracts nearest neighbors; - finds mutual nearest neighbors if required; - finally calls AdaLAM filtering. Inputs: k1: keypoint locations in the source image, in pixel coordinates. Expected an array with shape (num_keypoints_in_source_image, 2). k2: keypoint locations in the destination image, in pixel coordinates. Expected an array with shape (num_keypoints_in_destination_image, 2). d1: descriptors in the source image. Expected an array with shape (num_keypoints_in_source_image, descriptor_size). d2: descriptors in the destination image. Expected an array with shape (num_keypoints_in_destination_image, descriptor_size). im1shape: Shape of the source image. If None, it is inferred from keypoints max and min, at the cost of wasted runtime. So please provide it. Expected a tuple with (width, height) or (height, width) of source image im2shape: Shape of the destination image. If None, it is inferred from keypoints max and min, at the cost of wasted runtime. So please provide it. Expected a tuple with (width, height) or (height, width) of destination image o1/o2: keypoint orientations in degrees. They can be None if 'orientation_difference_threshold' in config is set to None. See documentation on 'orientation_difference_threshold' in the DEFAULT_CONFIG. Expected an array with shape (num_keypoints_in_source/destination_image,) s1/s2: keypoint scales. They can be None if 'scale_rate_threshold' in config is set to None. See documentation on 'scale_rate_threshold' in the DEFAULT_CONFIG. Expected an array with shape (num_keypoints_in_source/destination_image,) Returns: Filtered putative matches. A long tensor with shape (num_filtered_matches, 2) with indices of corresponding keypoints in k1 and k2. """ # noqa: E501 if s1 is None or s2 is None: if self.config['scale_rate_threshold'] is not None: raise AttributeError( "Current configuration considers keypoint scales for filtering, but scales have not been provided.\n" # noqa: E501 "Please either provide scales or set 'scale_rate_threshold' to None to disable scale filtering" ) if o1 is None or o2 is None: if self.config['orientation_difference_threshold'] is not None: raise AttributeError( "Current configuration considers keypoint orientations for filtering, but orientations have not been provided.\n" # noqa: E501 "Please either provide orientations or set 'orientation_difference_threshold' to None to disable orientations filtering" # noqa: E501 ) k1, k2, d1, d2, o1, o2, s1, s2 = self.__to_torch(k1, k2, d1, d2, o1, o2, s1, s2) distmat = dist_matrix(d1, d2, is_normalized=False) dd12, nn12 = torch.topk(distmat, k=2, dim=1, largest=False) # (n1, 2) putative_matches = nn12[:, 0] scores = dd12[:, 0] / dd12[:, 1].clamp_min_(1e-3) if self.config['force_seed_mnn']: dd21, nn21 = torch.min(distmat, dim=0) # (n2,) mnn = nn21[putative_matches] == torch.arange(k1.shape[0], device=self.config['device']) else: mnn = None return self.filter_matches(k1, k2, putative_matches, scores, mnn, im1shape, im2shape, o1, o2, s1, s2) def __to_torch(self, *args): return ( a if a is None or torch.is_tensor(a) else torch.tensor(a, device=self.config['device'], dtype=torch.float32) for a in args )