3d-shape-retrieval/distance.py at Python · alexanderapers/3d-shape-retrieval · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
import time
import csv
import numpy as np
#from numba import njit
from mesh import Mesh
from features_mesh import Features_Mesh
from shape_features_mesh import Shape_Features_Mesh
from scipy.stats import wasserstein_distance
from scipy.special import rel_entr


class Distance:
    def __init__(self, dataset_name, exclude_list):
        self.csv = "features/" + dataset_name + "_all_features_normalized.csv"
        self.features = self.csv_to_dict()
        self.exclude_list = exclude_list
        self.norm_info = np.load("norm_info.npy")
        self.n_bins = int((len(self.features["m1.ply"]) - 5) / 5)

        # edit this to tweak weights
        #self.area, self.compactness, self.AABB_volume, self.diameter, self.eccentricity
        elem_weights = np.array([10, 0, 0, 10, 10])
        a3_weights = np.repeat(5, self.n_bins)
        d1_weights = np.repeat(1, self.n_bins)
        d2_weights = np.repeat(3, self.n_bins)
        d3_weights = np.repeat(3, self.n_bins)
        d4_weights = np.repeat(3, self.n_bins)

        self.weights = self.normalize_weights(elem_weights, a3_weights, d1_weights, d2_weights, d3_weights, d4_weights)

        #print(self.features["m518.ply"])
        #print(self.features["m514.ply"])
        #print("euclidean:", self.distance("m518.ply", "m514.ply", self.euclidean))
        #print("EMD:", self.distance("m518.ply", "m514.ply", self.euclidean_EMD))
        #print("KL:", self.distance("m518.ply", "m514.ply", self.euclidean_KL))

        # print("weight: area", self.weights[0])
        # print("weight: compactness", self.weights[1])
        # print("weight: AABB_volume", self.weights[2])
        # print("weight: diameter", self.weights[3])
        # print("weight: eccentricity", self.weights[4])
        # print("weight: A3", sum(self.weights[5:35]))
        # print("weight: D1", sum(self.weights[35:65]))
        # print("weight: D2", sum(self.weights[65:95]))
        # print("weight: D3", sum(self.weights[95:125]))
        # print("weight: D4", sum(self.weights[125:155]))

        # compiling numba
        #self.manhatten(np.array([1.0]), np.array([1.0]))
        #self.euclidean(np.array([1.0]), np.array([1.0]))
        #self.cosine(np.array([1.0]), np.array([1.0]))

        # print this to see result of query
        #result = self.query("LabeledDB_new/Octopus/121.off", self.euclidean_EMD, k=10)
        #print(result)

        # for r, d in result:
        #     print(r, d)

    # Helper function so that weights don't need to add up to 1 at the input level
    def normalize_weights(*weight):
        all = np.concatenate(weight[1:])
        totalweight = np.sum(all)
        all = all / totalweight
        return all


    def query(self, mesh_file_path, metric, k=10):
        #start_time = time.perf_counter()
        query_mesh = self.meshify(mesh_file_path)
        query_features = self.extract_features_mesh(query_mesh)
        result = self.find_k_most_similar(query_features, metric, k)
        #print("--- % seconds ---" % (time.perf_counter() - start_time))
        return result


    def query_inside_db(self, mesh_file_path, metric, k=10):
        result = self.find_k_most_similar(self.features[mesh_file_path.split("/")[-1]], metric, k)
        return result


    def csv_to_dict(self):
        with open(self.csv, 'r') as read_obj:
            features_dict = dict()
            csv_reader = csv.reader(read_obj)
            next(csv_reader)
            for row in csv_reader:
                mesh_name = row[0]
                mesh_features = np.array(row[2:]).astype(float)
                features_dict[mesh_name] = mesh_features
            return features_dict


    def distance(self, mesh_name_1, mesh_name_2, metric):
        a = self.weights * self.features[mesh_name_1]
        b = self.weights * self.features[mesh_name_2]
        return metric(a, b)


    def meshify(self, query_mesh_file_path):
        mesh = Mesh(query_mesh_file_path)
        mesh.resample_mesh()
        mesh.normalize_mesh()
        mesh.save_thumbnail()
        return mesh


    def extract_features_mesh(self, query_mesh):
        elem_features = Features_Mesh(query_mesh).get_all_elementary_features()
        shape_features = Shape_Features_Mesh(query_mesh).get_all_shape_features()
        elem_features[2:] = (elem_features[2:] - self.norm_info[:, 0]) / self.norm_info[:, 1]
        return np.array(elem_features[2:] + shape_features[1:])


    def find_k_most_similar(self, query_features, metric, k=10):
        distances = {x: 0 for x in self.features}
        for mesh_name in self.features:
            distances[mesh_name] = metric(self.weights * query_features, self.weights * self.features[mesh_name])
        return sorted(distances.items(), key=lambda item: item[1])[:k]


    @staticmethod
    #@njit()
    def euclidean(mesh_features_1, mesh_features_2):
        return np.linalg.norm(mesh_features_1 - mesh_features_2)


    @staticmethod
    #@njit()
    def manhatten(mesh_features_1, mesh_features_2):
        return np.sum(np.abs(mesh_features_1 - mesh_features_2))


    @staticmethod
    #@njit()
    def cosine(mesh_features_1, mesh_features_2):
        norm_1 = np.linalg.norm(mesh_features_1)
        norm_2 = np.linalg.norm(mesh_features_2)
        return 1 - (np.dot(mesh_features_1, mesh_features_2) / (norm_1 * norm_2))


    #@staticmethod
    #@njit()
    def euclidean_EMD(self, mesh_features_1, mesh_features_2):
        distances = np.zeros(6)
        elem_distance = Distance.euclidean(mesh_features_1[:5], mesh_features_2[:5])
        distances[0] = elem_distance
        # print("elem distance", elem_distance)
        j = 1
        for i in range(5, 5*self.n_bins + 5, self.n_bins):
            hist_distance = wasserstein_distance(np.arange(self.n_bins), np.arange(self.n_bins),
                mesh_features_1[i:i+self.n_bins], mesh_features_2[i:i+self.n_bins])
            distances[j] = hist_distance / self.n_bins
            #print("hist distance:", distances[j], j)
            j += 1
        return np.mean(distances)


    #@staticmethod
    #njit()
    # KL divergence is kinda broken since it cannot deal with 0 values
    def euclidean_KL(self, mesh_features_1, mesh_features_2):
        distances = np.zeros(6)
        elem_distance = Distance.euclidean(mesh_features_1[:5], mesh_features_2[:5])
        distances[0] = elem_distance
        j = 1
        for i in range(5, 5*self.n_bins + 5, self.n_bins):
            hist_distance = Distance.symm_KL(mesh_features_1[i:i+self.n_bins], mesh_features_2[i:i+self.n_bins])
            distances[j] = hist_distance
            j += 1
        return np.mean(distances)


    @staticmethod
    #njit()
    def symm_KL(mesh_features_1, mesh_features_2):
        return np.sum(rel_entr(mesh_features_1, mesh_features_2) + rel_entr(mesh_features_2, mesh_features_1))


    def mahalanobis(self, mesh_features_1, mesh_features_2):
        if not hasattr(self, 'inv_covariance'):
            all_vectors = []
            for mesh_name in self.features:
                all_vectors.append(self.features[mesh_name])

            cov = np.cov(np.stack(all_vectors).T)
            self.inv_covariance = np.linalg.pinv(cov)

        diff = mesh_features_1 - mesh_features_2
        return (diff @ self.inv_covariance @ diff) ** 0.5