Unverified Commit fd2c93c7 authored by Kelvin Rodriguez's avatar Kelvin Rodriguez Committed by GitHub
Browse files

Added modified change detection algorithm (#456) 

* minor bug fixes

* added ok and okm methods

* renamed okbm method

* removed unused imports

* better docs, np array inputs

* updated doc string to make types more explicit
parent 602540a1

autocnet/cg/change_detection.py

+155 −9
Original line number Diff line number Diff line
@@ -2,9 +2,9 @@ import numpy as np 
from matplotlib.path import Path

from shapely.geometry import Point, MultiPoint

import geopandas as gpd



import cv2

from sklearn.cluster import  OPTICS

from plio.io.io_gdal import GeoDataset



from autocnet.utils.utils import bytescale

from autocnet.matcher.cpu_extractor import extract_features
@@ -18,16 +18,161 @@ def image_diff(arr1, arr2): 
     diff = arr1-arr2

     diff[np.isnan(diff)] = 0



     return bytescale(diff)

     return diff





def okubogar_detector(image1, image2, nbins=50, extractor_method="orb", image_func=image_diff,

                      extractor_kwargs={"nfeatures": 2000, "scaleFactor": 1.1, "nlevels": 1}):

     """

     Simple change detection algorithm which produces an overlay image of change hotspots

     (i.e. a 2d histogram image of detected change density).



     Largely based on a method created by Chris Okubo and Brendon Bogar. Histogram step

     was added for readability.







     image1

           \

             image subtraction/ratio -> feature_extraction -> feature_histogram

           /

     image2



     TODO: Paper/abstract might exist, cite



     Parameters

     ----------



     image1 : np.array, plio.GeoDataset

             Image representing the "before" state of the ROI, can be a 2D numpy array or plio GeoDataset



     image2 : np.array, plio.GeoDataset

             Image representing the "after" state of the ROI, can be a 2D numpy array or plio GeoDataset



     image_func : callable

                  Function used to create a derived image from image1 and image2, which in turn is

                  the input for the feature extractor. The first two arguments are 2d numpy arrays, image1 and image2,

                  and must return a 2d numpy array. Default function returns a difference image.



                  Example func:



                  >>> from numpy import random

                  >>> image1, image2 = random.random((50,50)), random.random((50,50))

                  >>> def ratio(image1, image2):

                  >>>   # do stuff with image1 and image2

                  >>>   new_image = image1/image2

                  >>>   return new_image # must return a single variable, a 2D numpy array

                  >>> results = okubogar_detector(image1, image2, image_func=ratio)



def okubogar_detector(image1, image2, nbins=50, extractor_method="orb", extractor_kwargs={"nfeatures": 2000, "scaleFactor": 1.1, "nlevels": 1}, cluster_params={"min_samples": 10, "max_eps": 10, "eps": .5, "xi":.5},  image_func=image_diff):

     arr1 = image1.read_array()

     arr2 = image2.read_array()

     arr1[arr1 == arr1.min()] = 0

     arr2[arr2 == arr2.min()] = 0

     arr1 = bytescale(arr1)

     arr2 = bytescale(arr2)

                  Or, alternatively:



                  >>> from numpy import random

                  >>> image1, image2 = random.random((50,50)), random.random((50,50))

                  >>> results = okubogar_detector(image1, image2, image_func=lambda im1, im2: im1/im2)



     nbins : int

            number of bins to use in the 2d histogram



     extractor_method : {'orb', 'sift', 'fast', 'surf', 'vl_sift'}

               The detector method to be used.  Note that vl_sift requires that

               vlfeat and cyvlfeat dependencies be installed.



     extractor_kwargs : dict

                        A dictionary containing OpenCV SIFT parameters names and values.



     See Also

     --------



     feature extractor: autocnet.matcher.cpu_extractor.extract_features



     """

     if isinstance(image1, GeoDataset):

         image1 = image1.read_array()



     if isinstance(image2, GeoDataset):

         image2 = image2.read_array()



     image1[image1 == image1.min()] = 0

     image2[image2 == image2.min()] = 0

     arr1 = bytescale(image1)

     arr2 = bytescale(image2)



     bdiff = image_func(arr1, arr2)



     keys, descriptors = extract_features(bdiff, extractor_method, extractor_parameters=extractor_kwargs)

     x,y = keys["x"], keys["y"]



     points = [Point(xval, yval) for xval,yval in zip(x,y)]



     heatmap, xedges, yedges = np.histogram2d(y, x, bins=nbins, range=[[0, bdiff.shape[0]], [0, bdiff.shape[1]]])

     heatmap = cv2.resize(heatmap, dsize=(bdiff.shape[1], bdiff.shape[0]), interpolation=cv2.INTER_NEAREST)



     return points, heatmap, bdiff





def okbm_detector(image1, image2, nbins=50, extractor_method="orb",  image_func=image_diff,

                 extractor_kwargs={"nfeatures": 2000, "scaleFactor": 1.1, "nlevels": 1},

                 cluster_params={"min_samples": 10, "max_eps": 10, "eps": .5, "xi":.5}):

     """

     okobubogar modified detector, experimental feature based change detection algorithmthat expands on okobubogar to allow for

     programmatic change detection. Returns detected feature changes as weighted polygons.





     Parameters

     ----------



     image1 : GeoDataset

             Image representing the "before" state of the ROI, can be a 2D numpy array or plio GeoDataset



     image2 : GeoDataset

             Image representing the "after" state of the ROI, can be a 2D numpy array or plio GeoDataset



     image_func : callable

                  Function used to create a derived image from image1 and image2, which in turn is

                  the input for the feature extractor. The first two arguments are 2d numpy arrays, image1 and image2,

                  and must return a 2d numpy array. Default function returns a difference image.



                  Example func:



                  >>> from numpy import random

                  >>> image1, image2 = random.random((50,50)), random.random((50,50))

                  >>> def ratio(image1, image2):

                  >>>   # do stuff with image1 and image2

                  >>>   new_image = image1/image2

                  >>>   return new_image # must return a single variable, a 2D numpy array

                  >>> results = okbm_detector(image1, image2, image_func=ratio)



                  Or, alternatively:



                  >>> from numpy import random

                  >>> image1, image2 = random.random((50,50)), random.random((50,50))

                  >>> results = okbm_detector(image1, image2, image_func=lambda im1, im2: im1/im2)



     nbins : int

            number of bins to use in the 2d histogram



     extractor_method : {'orb', 'sift', 'fast', 'surf', 'vl_sift'}

               The detector method to be used.  Note that vl_sift requires that

               vlfeat and cyvlfeat dependencies be installed.



     extractor_kwargs : dict

                        A dictionary containing OpenCV SIFT parameters names and values.



     cluster_params : dict

                      A dictionary containing sklearn.cluster.OPTICS parameters



     """



     if isinstance(image1, GeoDataset):

         image1 = image1.read_array()



     if isinstance(image2, GeoDataset):

         image2 = image2.read_array()



     image1[image1 == image1.min()] = 0

     image2[image2 == image2.min()] = 0

     arr1 = bytescale(image1)

     arr2 = bytescale(image2)



     bdiff = image_func(arr1, arr2)
@@ -70,3 +215,4 @@ def okubogar_detector(image1, image2, nbins=50, extractor_method="orb", extracto 


     return polys, weights, bdiff

autocnet/graph/edge.py

+2 −1
Original line number Diff line number Diff line
@@ -343,6 +343,7 @@ class Edge(dict, MutableMapping): 
        _, mask = self.clean(clean_keys)

        s_keypoints, d_keypoints = self.get_match_coordinates(clean_keys=clean_keys)

        self.fundamental_matrix, fmask = fm.compute_fundamental_matrix(s_keypoints, d_keypoints, **kwargs)

        fmask = fmask.flatten()



        if isinstance(self.fundamental_matrix, np.ndarray):

            # Convert the truncated RANSAC mask back into a full length mask