-
Notifications
You must be signed in to change notification settings - Fork 24
/
llff.py
144 lines (125 loc) · 5.65 KB
/
llff.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 numpy as np
import imageio
import torch
from torch.utils.data import Dataset
import sys
sys.path.append("../")
from .data_utils import random_crop, random_flip, get_nearest_pose_ids
from .llff_data_utils import load_llff_data, batch_parse_llff_poses
class LLFFDataset(Dataset):
def __init__(self, args, mode, **kwargs):
base_dir = os.path.join(args.rootdir, "data/real_iconic_noface/")
self.args = args
self.mode = mode # train / test / validation
self.num_source_views = args.num_source_views
self.render_rgb_files = []
self.render_intrinsics = []
self.render_poses = []
self.render_train_set_ids = []
self.render_depth_range = []
self.train_intrinsics = []
self.train_poses = []
self.train_rgb_files = []
scenes = os.listdir(base_dir)
for i, scene in enumerate(scenes):
scene_path = os.path.join(base_dir, scene)
_, poses, bds, render_poses, i_test, rgb_files = load_llff_data(
scene_path, load_imgs=False, factor=4
)
near_depth = np.min(bds)
far_depth = np.max(bds)
intrinsics, c2w_mats = batch_parse_llff_poses(poses)
if mode == "train":
i_train = np.array(np.arange(int(poses.shape[0])))
i_render = i_train
else:
i_test = np.arange(poses.shape[0])[:: self.args.llffhold]
i_train = np.array(
[
j
for j in np.arange(int(poses.shape[0]))
if (j not in i_test and j not in i_test)
]
)
i_render = i_test
self.train_intrinsics.append(intrinsics[i_train])
self.train_poses.append(c2w_mats[i_train])
self.train_rgb_files.append(np.array(rgb_files)[i_train].tolist())
num_render = len(i_render)
self.render_rgb_files.extend(np.array(rgb_files)[i_render].tolist())
self.render_intrinsics.extend([intrinsics_ for intrinsics_ in intrinsics[i_render]])
self.render_poses.extend([c2w_mat for c2w_mat in c2w_mats[i_render]])
self.render_depth_range.extend([[near_depth, far_depth]] * num_render)
self.render_train_set_ids.extend([i] * num_render)
def __len__(self):
return len(self.render_rgb_files)
def __getitem__(self, idx):
rgb_file = self.render_rgb_files[idx]
rgb = imageio.imread(rgb_file).astype(np.float32) / 255.0
render_pose = self.render_poses[idx]
intrinsics = self.render_intrinsics[idx]
depth_range = self.render_depth_range[idx]
train_set_id = self.render_train_set_ids[idx]
train_rgb_files = self.train_rgb_files[train_set_id]
train_poses = self.train_poses[train_set_id]
train_intrinsics = self.train_intrinsics[train_set_id]
img_size = rgb.shape[:2]
camera = np.concatenate(
(list(img_size), intrinsics.flatten(), render_pose.flatten())
).astype(np.float32)
if self.mode == "train":
id_render = train_rgb_files.index(rgb_file)
subsample_factor = np.random.choice(np.arange(1, 4), p=[0.2, 0.45, 0.35])
num_select = self.num_source_views + np.random.randint(low=-2, high=3)
else:
id_render = -1
subsample_factor = 1
num_select = self.num_source_views
nearest_pose_ids = get_nearest_pose_ids(
render_pose,
train_poses,
min(self.num_source_views * subsample_factor, 20),
tar_id=id_render,
angular_dist_method="dist",
)
nearest_pose_ids = np.random.choice(
nearest_pose_ids, min(num_select, len(nearest_pose_ids)), replace=False
)
assert id_render not in nearest_pose_ids
# occasionally include input image
if np.random.choice([0, 1], p=[0.995, 0.005]) and self.mode == "train":
nearest_pose_ids[np.random.choice(len(nearest_pose_ids))] = id_render
src_rgbs = []
src_cameras = []
for id in nearest_pose_ids:
src_rgb = imageio.imread(train_rgb_files[id]).astype(np.float32) / 255.0
train_pose = train_poses[id]
train_intrinsics_ = train_intrinsics[id]
src_rgbs.append(src_rgb)
img_size = src_rgb.shape[:2]
src_camera = np.concatenate(
(list(img_size), train_intrinsics_.flatten(), train_pose.flatten())
).astype(np.float32)
src_cameras.append(src_camera)
src_rgbs = np.stack(src_rgbs, axis=0)
src_cameras = np.stack(src_cameras, axis=0)
if self.mode == "train":
crop_h = np.random.randint(low=250, high=750)
crop_h = crop_h + 1 if crop_h % 2 == 1 else crop_h
crop_w = int(400 * 600 / crop_h)
crop_w = crop_w + 1 if crop_w % 2 == 1 else crop_w
rgb, camera, src_rgbs, src_cameras = random_crop(
rgb, camera, src_rgbs, src_cameras, (crop_h, crop_w)
)
if self.mode == "train" and np.random.choice([0, 1]):
rgb, camera, src_rgbs, src_cameras = random_flip(rgb, camera, src_rgbs, src_cameras)
depth_range = torch.tensor([depth_range[0] * 0.9, depth_range[1] * 1.6])
return {
"rgb": torch.from_numpy(rgb[..., :3]),
"camera": torch.from_numpy(camera),
"rgb_path": rgb_file,
"src_rgbs": torch.from_numpy(src_rgbs[..., :3]),
"src_cameras": torch.from_numpy(src_cameras),
"depth_range": depth_range,
}