-
Notifications
You must be signed in to change notification settings - Fork 193
/
DVRDataset.py
274 lines (241 loc) · 9.54 KB
/
DVRDataset.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
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
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
import os
import torch
import torch.nn.functional as F
import glob
import imageio
import numpy as np
import cv2
from util import get_image_to_tensor_balanced, get_mask_to_tensor
class DVRDataset(torch.utils.data.Dataset):
"""
Dataset from DVR (Niemeyer et al. 2020)
Provides 3D-R2N2 and NMR renderings
"""
def __init__(
self,
path,
stage="train",
list_prefix="softras_",
image_size=None,
sub_format="shapenet",
scale_focal=True,
max_imgs=100000,
z_near=1.2,
z_far=4.0,
skip_step=None,
):
"""
:param path dataset root path, contains metadata.yml
:param stage train | val | test
:param list_prefix prefix for split lists: <list_prefix>[train, val, test].lst
:param image_size result image size (resizes if different); None to keep original size
:param sub_format shapenet | dtu dataset sub-type.
:param scale_focal if true, assume focal length is specified for
image of side length 2 instead of actual image size. This is used
where image coordinates are placed in [-1, 1].
"""
super().__init__()
self.base_path = path
assert os.path.exists(self.base_path)
cats = [x for x in glob.glob(os.path.join(path, "*")) if os.path.isdir(x)]
if stage == "train":
file_lists = [os.path.join(x, list_prefix + "train.lst") for x in cats]
elif stage == "val":
file_lists = [os.path.join(x, list_prefix + "val.lst") for x in cats]
elif stage == "test":
file_lists = [os.path.join(x, list_prefix + "test.lst") for x in cats]
all_objs = []
for file_list in file_lists:
if not os.path.exists(file_list):
continue
base_dir = os.path.dirname(file_list)
cat = os.path.basename(base_dir)
with open(file_list, "r") as f:
objs = [(cat, os.path.join(base_dir, x.strip())) for x in f.readlines()]
all_objs.extend(objs)
self.all_objs = all_objs
self.stage = stage
self.image_to_tensor = get_image_to_tensor_balanced()
self.mask_to_tensor = get_mask_to_tensor()
print(
"Loading DVR dataset",
self.base_path,
"stage",
stage,
len(self.all_objs),
"objs",
"type:",
sub_format,
)
self.image_size = image_size
if sub_format == "dtu":
self._coord_trans_world = torch.tensor(
[[1, 0, 0, 0], [0, -1, 0, 0], [0, 0, -1, 0], [0, 0, 0, 1]],
dtype=torch.float32,
)
self._coord_trans_cam = torch.tensor(
[[1, 0, 0, 0], [0, -1, 0, 0], [0, 0, -1, 0], [0, 0, 0, 1]],
dtype=torch.float32,
)
else:
self._coord_trans_world = torch.tensor(
[[1, 0, 0, 0], [0, 0, -1, 0], [0, 1, 0, 0], [0, 0, 0, 1]],
dtype=torch.float32,
)
self._coord_trans_cam = torch.tensor(
[[1, 0, 0, 0], [0, -1, 0, 0], [0, 0, -1, 0], [0, 0, 0, 1]],
dtype=torch.float32,
)
self.sub_format = sub_format
self.scale_focal = scale_focal
self.max_imgs = max_imgs
self.z_near = z_near
self.z_far = z_far
self.lindisp = False
def __len__(self):
return len(self.all_objs)
def __getitem__(self, index):
cat, root_dir = self.all_objs[index]
rgb_paths = [
x
for x in glob.glob(os.path.join(root_dir, "image", "*"))
if (x.endswith(".jpg") or x.endswith(".png"))
]
rgb_paths = sorted(rgb_paths)
mask_paths = sorted(glob.glob(os.path.join(root_dir, "mask", "*.png")))
if len(mask_paths) == 0:
mask_paths = [None] * len(rgb_paths)
if len(rgb_paths) <= self.max_imgs:
sel_indices = np.arange(len(rgb_paths))
else:
sel_indices = np.random.choice(len(rgb_paths), self.max_imgs, replace=False)
rgb_paths = [rgb_paths[i] for i in sel_indices]
mask_paths = [mask_paths[i] for i in sel_indices]
cam_path = os.path.join(root_dir, "cameras.npz")
all_cam = np.load(cam_path)
all_imgs = []
all_poses = []
all_masks = []
all_bboxes = []
focal = None
if self.sub_format != "shapenet":
# Prepare to average intrinsics over images
fx, fy, cx, cy = 0.0, 0.0, 0.0, 0.0
for idx, (rgb_path, mask_path) in enumerate(zip(rgb_paths, mask_paths)):
i = sel_indices[idx]
img = imageio.imread(rgb_path)[..., :3]
if self.scale_focal:
x_scale = img.shape[1] / 2.0
y_scale = img.shape[0] / 2.0
xy_delta = 1.0
else:
x_scale = y_scale = 1.0
xy_delta = 0.0
if mask_path is not None:
mask = imageio.imread(mask_path)
if len(mask.shape) == 2:
mask = mask[..., None]
mask = mask[..., :1]
if self.sub_format == "dtu":
# Decompose projection matrix
# DVR uses slightly different format for DTU set
P = all_cam["world_mat_" + str(i)]
P = P[:3]
K, R, t = cv2.decomposeProjectionMatrix(P)[:3]
K = K / K[2, 2]
pose = np.eye(4, dtype=np.float32)
pose[:3, :3] = R.transpose()
pose[:3, 3] = (t[:3] / t[3])[:, 0]
scale_mtx = all_cam.get("scale_mat_" + str(i))
if scale_mtx is not None:
norm_trans = scale_mtx[:3, 3:]
norm_scale = np.diagonal(scale_mtx[:3, :3])[..., None]
pose[:3, 3:] -= norm_trans
pose[:3, 3:] /= norm_scale
fx += torch.tensor(K[0, 0]) * x_scale
fy += torch.tensor(K[1, 1]) * y_scale
cx += (torch.tensor(K[0, 2]) + xy_delta) * x_scale
cy += (torch.tensor(K[1, 2]) + xy_delta) * y_scale
else:
# ShapeNet
wmat_inv_key = "world_mat_inv_" + str(i)
wmat_key = "world_mat_" + str(i)
if wmat_inv_key in all_cam:
extr_inv_mtx = all_cam[wmat_inv_key]
else:
extr_inv_mtx = all_cam[wmat_key]
if extr_inv_mtx.shape[0] == 3:
extr_inv_mtx = np.vstack((extr_inv_mtx, np.array([0, 0, 0, 1])))
extr_inv_mtx = np.linalg.inv(extr_inv_mtx)
intr_mtx = all_cam["camera_mat_" + str(i)]
fx, fy = intr_mtx[0, 0], intr_mtx[1, 1]
assert abs(fx - fy) < 1e-9
fx = fx * x_scale
if focal is None:
focal = fx
else:
assert abs(fx - focal) < 1e-5
pose = extr_inv_mtx
pose = (
self._coord_trans_world
@ torch.tensor(pose, dtype=torch.float32)
@ self._coord_trans_cam
)
img_tensor = self.image_to_tensor(img)
if mask_path is not None:
mask_tensor = self.mask_to_tensor(mask)
rows = np.any(mask, axis=1)
cols = np.any(mask, axis=0)
rnz = np.where(rows)[0]
cnz = np.where(cols)[0]
if len(rnz) == 0:
raise RuntimeError(
"ERROR: Bad image at", rgb_path, "please investigate!"
)
rmin, rmax = rnz[[0, -1]]
cmin, cmax = cnz[[0, -1]]
bbox = torch.tensor([cmin, rmin, cmax, rmax], dtype=torch.float32)
all_masks.append(mask_tensor)
all_bboxes.append(bbox)
all_imgs.append(img_tensor)
all_poses.append(pose)
if self.sub_format != "shapenet":
fx /= len(rgb_paths)
fy /= len(rgb_paths)
cx /= len(rgb_paths)
cy /= len(rgb_paths)
focal = torch.tensor((fx, fy), dtype=torch.float32)
c = torch.tensor((cx, cy), dtype=torch.float32)
all_bboxes = None
elif mask_path is not None:
all_bboxes = torch.stack(all_bboxes)
all_imgs = torch.stack(all_imgs)
all_poses = torch.stack(all_poses)
if len(all_masks) > 0:
all_masks = torch.stack(all_masks)
else:
all_masks = None
if self.image_size is not None and all_imgs.shape[-2:] != self.image_size:
scale = self.image_size[0] / all_imgs.shape[-2]
focal *= scale
if self.sub_format != "shapenet":
c *= scale
elif mask_path is not None:
all_bboxes *= scale
all_imgs = F.interpolate(all_imgs, size=self.image_size, mode="area")
if all_masks is not None:
all_masks = F.interpolate(all_masks, size=self.image_size, mode="area")
result = {
"path": root_dir,
"img_id": index,
"focal": focal,
"images": all_imgs,
"poses": all_poses,
}
if all_masks is not None:
result["masks"] = all_masks
if self.sub_format != "shapenet":
result["c"] = c
else:
result["bbox"] = all_bboxes
return result