-
Notifications
You must be signed in to change notification settings - Fork 8
/
datasets.py
144 lines (115 loc) · 5 KB
/
datasets.py
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
import os
import torch
from torch.utils.data import Dataset
import imageio
import numpy as np
from utils import get_mgrid
def rescale_img(img, tmin=-1.0, tmax=1.0):
img = np.clip(img, tmin, tmax)
img = ((img + 1.0) * 255.0).astype(np.uint8)
return img
def sample_points(coords, normals, on_surface_samples):
point_cloud_size = coords.shape[0]
on_surface_samples = min(on_surface_samples, point_cloud_size)
off_surface_samples = on_surface_samples
# random coords on the surface
rand_idx = np.random.choice(point_cloud_size, size=on_surface_samples)
on_surface_coords = coords[rand_idx, :]
on_surface_normals = normals[rand_idx, :]
on_surface_sdfs = np.zeros((on_surface_samples, 1))
# random coords in the volume
off_surface_coords = np.random.uniform(-1, 1, (off_surface_samples, 3))
off_surface_normals = np.ones((off_surface_samples, 3)) * -1
off_surface_sdfs = np.ones((off_surface_samples, 1)) * -1
# outputs
coords = np.concatenate((on_surface_coords, off_surface_coords), axis=0)
sdf = np.concatenate((on_surface_sdfs, off_surface_sdfs), axis=0)
normals = np.concatenate((on_surface_normals, off_surface_normals), axis=0)
return coords.astype(np.float32), sdf.astype(np.float32), normals.astype(np.float32)
# Reshape point cloud such that it lies in bounding box of (-1, 1) * scale
def normalize_points(coords, scale=0.9):
coords -= np.mean(coords, axis=0, keepdims=True)
coord_max, coord_min = np.amax(coords), np.amin(coords)
coords = (coords - coord_min) / (coord_max - coord_min) # (0, 1)
coords = (coords - 0.5) * (2.0 * scale)
return coords
def load_points(filename: str):
if filename.endswith('.xyz'):
points = np.loadtxt(filename)
elif filename.endswith('.npy'):
points = np.load(filename)
else:
raise NotImplementedError
return points
class PointCloud(Dataset):
def __init__(self, pointcloud_path, on_surface_samples, scale=0.9,
normalize=True, **kwargs):
super().__init__()
point_cloud = load_points(pointcloud_path)
self.on_surface_samples = on_surface_samples
assert on_surface_samples < point_cloud.shape[0]
self.coords = point_cloud[:, :3]
self.normals = point_cloud[:, 3:]
if normalize:
self.coords = normalize_points(self.coords, scale)
def __len__(self):
return self.coords.shape[0] // self.on_surface_samples
def __getitem__(self, idx):
coords, sdfs, normals = sample_points(self.coords, self.normals, self.on_surface_samples)
return torch.from_numpy(coords), torch.from_numpy(sdfs), torch.from_numpy(normals)
class DFaustDataset(Dataset):
def __init__(self, root_folder, filename, pc_num, in_memory=False, **kwargs):
super().__init__()
self.root_folder = root_folder
self.filenames = self.get_filenames(filename)
self.on_surface_samples = pc_num
self.in_memory = in_memory
self.points = [None] * len(self.filenames)
if in_memory:
print('Load {} files into memory.'.format(len(self.filenames)))
for i, filename in enumerate(self.filenames):
self.points[i] = load_points(filename)
def get_filenames(self, filelist):
with open(filelist, 'r') as fid:
lines = fid.readlines()
filenames = [os.path.join(self.root_folder, line.strip()) for line in lines]
return filenames
def __len__(self):
return len(self.filenames)
def __getitem__(self, idx):
if self.in_memory:
points = self.points[idx]
else:
filename = self.filenames[idx]
points = load_points(filename)
coords, sdfs, normals = sample_points(points[:, :3], points[:, 3:], self.on_surface_samples)
return torch.from_numpy(coords), torch.from_numpy(sdfs), torch.from_numpy(normals), idx
class SingleImage(Dataset):
def __init__(self, filename):
img = np.asarray(imageio.imread(filename)).astype(np.float32)
img = img / 255.0 # [0, 1]
# img = img * (2.0 / 255.0) - 1.0 # [-1, 1]
assert img.shape[0] == img.shape[1]
self.img = img.reshape(1, img.shape[0]*img.shape[1], -1)
self.channel = img.shape[-1]
self.resolution = img.shape[0]
coords = get_mgrid(self.resolution, dim=2, offset=0.5, r=-1) # [-1, 1]
self.coords = np.expand_dims(coords, axis=0)
def __len__(self):
return 1
def __getitem__(self, idx):
return torch.from_numpy(self.coords), torch.from_numpy(self.img)
class SDFVolume(Dataset):
def __init__(self, filename, pc_num):
self.pc_num = pc_num
sdf = np.load(filename)
assert sdf.shape[0] == sdf.shape[1] and sdf.shape[1] == sdf.shape[2]
self.resolution = sdf.shape[0]
self.total_num = sdf.shape[0] ** 3
self.sdf = sdf.reshape(-1, 1)
self.coords = get_mgrid(self.resolution, dim=3, offset=0, r=1)
def __len__(self):
return self.total_num // self.pc_num
def __getitem__(self, idx):
rnd = np.random.choice(self.total_num, self.pc_num)
return torch.from_numpy(self.coords[rnd, :]), torch.from_numpy(self.sdf[rnd, :])