CameraAwareBodyModel/data.py at main · houchenst/CameraAwareBodyModel · 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
191
192
'''
2D and 3D occupancy datasets
'''

import torch
from torch.utils.data import Dataset
import random
import numpy as np
from utils import rotation_matrix, random_cam_center, angle_features, uv_features
from tqdm import tqdm

class OccDataset3D(Dataset):

    def __init__(self, obj, bbox, num_samples, noise=0.03):
        '''
        :param obj: An object that has an occupancy function and a surface sampling function
        :param bbox: A tuple of tuples with the max and min value of each dimension (x,y,z)
        :param num_samples: The number of samples to take from the surface of the object. Twice this
                            number of samples will be taken uniformly from within the bbox
        '''
        # store inputs
        self.obj = obj
        ((self.x_min, self.x_max),(self.y_min, self.y_max),(self.z_min, self.z_max)) = bbox
        self.num_samples = num_samples

        # random uniform sampling within bbox
        self.points = [[random.uniform(self.x_min, self.x_max), random.uniform(self.y_min, self.y_max), random.uniform(self.z_min, self.z_max)] for _ in range(2*num_samples)]
        # surface sampling
        self.points += self.obj.sample_surface(num_samples, gaussian_noise=noise)

        # getting point labels
        self.labels = [torch.tensor(self.obj.contains(x)) for x in self.points]

        # make points tensors
        self.points = [torch.tensor(point) for point in self.points]

        self.length = len(self.points)


    def __len__(self):
        '''
        Returns the length of the dataset
        '''
        return self.length

    def __getitem__(self, idx):
        '''
        Gets a tuple of (point, label) at given index
        '''
        return self.points[idx], self.labels[idx]


class OccDataset2D(Dataset):

    def __init__(self, obj, num_samples, focal_length, sensor_bounds, sensor_size, cam_radius, ray_samples=20, min_cameras=30):
        '''
        :param obj: An object that has an occupancy function and a surface sampling function
        :param num_samples: The number of individual 2D occupancy points to generate for training
        :param sensor_bounds: The bounds of the sensor in 2d camera coordinates. [a,b] will select the
                              box ranging from -a to a in x and -b to b in y
        :param sensor_size: The resolution of 2d occupancy images that we generate (wxh in pixels)
        :param focal_length: The focal length to use for the camera when making 2D occupancy images
        :param cam_radius: The radius of the sphere from which cam centers should be sampled
        :param ray_samples: The number of 3D occupancy queries to make along a ray to see if the 2D point is occupied
        :param min_cameras: The minimum number of camera locations to include data for.
                            Pass 1 to force using every pixel for each camera
        '''
        # store inputs
        self.obj = obj
        self.num_samples = num_samples
        self.focal_length = focal_length
        self.sensor_bounds = sensor_bounds
        self.sensor_size = sensor_size
        self.cam_radius = cam_radius
        self.ray_samples = ray_samples
        # self.pos_embedding_size = pos_embedding_size

        self.data = []
        self.view_embeddings = []
        self.labels = []

        # we want a variety of camera views, so if the sensor size is large, we might not use every pixel
        self.samples_per_camera = int(min(self.sensor_size[0]*self.sensor_size[1], self.num_samples / min_cameras))

        print("Generating 2D Data...")
        pbar = tqdm(total=num_samples)
        while len(self.data) < self.num_samples:
            cam_center = random_cam_center(self.cam_radius)
            l, x, y, _, _ = obj.contains_2d(cam_center, self.focal_length, self.sensor_bounds, self.sensor_size, ray_samples=self.ray_samples)
            l = l.flatten()
            x = x.flatten()
            y = y.flatten()
            r = rotation_matrix(cam_center, inverse=True).flatten()
            # cam_params = list(r) + cam_center
            view_embedding = angle_features(cam_center)

            # samples some fraction of the pixels to add to our data
            for i in random.sample(range(l.shape[0]), self.samples_per_camera):
                self.labels.append(torch.tensor(l[i], dtype=torch.float32))
                # self.data.append(torch.tensor(uv_features([x[i],y[i]], self.sensor_size[0], self.sensor_size[1]), dtype=torch.float32))
                self.data.append(torch.tensor([x[i], y[i]], dtype=torch.float32))
                # self.cam_params.append(torch.tensor(cam_params, dtype=torch.float32))
                self.view_embeddings.append(torch.tensor(view_embedding, dtype=torch.float32))
            pbar.update(self.samples_per_camera)
        tqdm._instances.clear()

        self.length = len(self.data)


    def __len__(self):
        '''
        Returns the length of the dataset
        '''
        return self.length

    def __getitem__(self, idx):
        '''
        Gets a tuple of (point, label) at given index
        '''
        return self.data[idx], self.labels[idx], self.view_embeddings[idx]

class OccDataset3DUVD(Dataset):

    def __init__(self, obj, num_samples, focal_length, sensor_bounds, sensor_size, cam_radius, ray_samples=20, min_cameras=30):
        '''
        :param obj: An object that has an occupancy function and a surface sampling function
        :param num_samples: The number of individual 2D occupancy points to generate for training
        :param sensor_bounds: The bounds of the sensor in 2d camera coordinates. [a,b] will select the
                              box ranging from -a to a in x and -b to b in y
        :param sensor_size: The resolution of 2d occupancy images that we generate (wxh in pixels)
        :param focal_length: The focal length to use for the camera when making 2D occupancy images
        :param cam_radius: The radius of the sphere from which cam centers should be sampled
        :param ray_samples: The number of 3D occupancy queries to make along a ray to see if the 2D point is occupied
        :param min_cameras: The minimum number of camera locations to include data for.
                            Pass 1 to force using every pixel for each camera
        '''
        # store inputs
        self.obj = obj
        self.num_samples = num_samples
        self.focal_length = focal_length
        self.sensor_bounds = sensor_bounds
        self.sensor_size = sensor_size
        self.cam_radius = cam_radius
        self.ray_samples = ray_samples
        self.pos_embedding_size = pos_embedding_size

        self.uv = []
        self.xyz = []
        self.cam_params = []
        self.labels = []

        # we want a variety of camera views, so if the sensor size is large, we might not use every pixel
        self.samples_per_camera = int(min(self.sensor_size[0]*self.sensor_size[1], self.num_samples / min_cameras))

        print("Generating 3D UVD Data...")
        pbar = tqdm(total=num_samples)
        while len(self.xyz) < self.num_samples:
            cam_center = random_cam_center(self.cam_radius)
            _, u, v, occ_3d, coords_3d = obj.contains_2d(cam_center, self.focal_length, self.sensor_bounds, self.sensor_size, ray_samples=self.ray_samples)
            coords_3d = np.reshape(coords_3d, (-1,3))
            occ_3d = occ_3d.flatten()
            u = np.stack([u]*5, axis=-1).flatten()
            v = np.stack([v]*5, axis=-1).flatten()
            r = rotation_matrix(cam_center, inverse=True).flatten()
            cam_params = list(r) + cam_center
            # view_embedding = angle_features(cam_center, L=self.pos_embedding_size)

            # samples some fraction of the pixels to add to our data
            for i in random.sample(range(coords_3d.shape[0]), self.samples_per_camera):
                self.labels.append(torch.tensor(occ_3d[i], dtype=torch.float32))
                self.xyz.append(torch.tensor(coords_3d[i], dtype=torch.float32))
                # self.uv.append(torch.tensor(uv_features([u[i],v[i]], self.sensor_size[0], self.sensor_size[1]), dtype=torch.float32))
                self.uv.append(torch.tensor([u[i], v[i]], dtype=torch.float32))
                self.cam_params.append(torch.tensor(cam_params, dtype=torch.float32))
                # self.view_embeddings.append(torch.tensor(view_embedding, dtype=torch.float32))
            pbar.update(self.samples_per_camera)
        tqdm._instances.clear()

        self.length = len(self.xyz)


    def __len__(self):
        '''
        Returns the length of the dataset
        '''
        return self.length

    def __getitem__(self, idx):
        '''
        Gets a tuple of (point, label) at given index
        '''
        return self.uv[idx], self.xyz[idx], self.labels[idx], self.cam_params[idx]