/
drunkards.py
executable file
·200 lines (173 loc) · 9 KB
/
drunkards.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
import numpy as np
import torch
import torch.utils.data as data
import torch.nn.functional as F
import os
import cv2
import math
import random
import os.path as osp
import flowpy
from glob import glob
import drunkards_odometry.projective_ops as pops
from .augmentation import ColorAugmentor
from drunkards_odometry.sampler_ops import bilinear_sampler
class DrunkDataset(data.Dataset):
def __init__(self, root, difficulty_level=None, do_augment=True, res_factor=1., scenes_to_use=None,
depth_augmentor=False, mode='train', invert_order_prob=0.0):
self.init_seed = None
self.res_factor = res_factor
self.depth_augmentor = depth_augmentor
self.mode = mode
self.invert_order_prob = invert_order_prob
if do_augment:
self.augmentor = ColorAugmentor()
else:
self.augmentor = None
self.image_list = []
self.depth_list = []
self.flow_list = []
self.pose_list = []
self.intrinsics_list = []
intrinsics = np.array([190.68059285, 286.02088928, 160., 160.]) # Drunkard's Dataset 320x320
available_scenes = [i.name.lstrip('0') or '0' for i in os.scandir(root) if i.is_dir()]
available_scenes = list(map(int, available_scenes))
if isinstance(scenes_to_use, int):
scenes_to_use = [scenes_to_use]
if isinstance(available_scenes, int):
available_scenes = [available_scenes]
scenes = list(set(scenes_to_use) & set(available_scenes))
for scene_i in scenes:
images = sorted(glob(osp.join(
root, "{:05d}".format(scene_i), "level{}/color/*.png".format(difficulty_level)
)))
if not len(images):
images = sorted(glob(osp.join(
root, "{:05d}".format(scene_i), "level{}/color/color/*.png".format(difficulty_level)
)))
depths = sorted(glob(osp.join(
root, "{:05d}".format(scene_i), "level{}/depth/*.png".format(difficulty_level)
)))
if not len(depths):
depths = sorted(glob(osp.join(
root, "{:05d}".format(scene_i), "level{}/depth/depth/*.png".format(difficulty_level)
)))
flows = sorted(glob(osp.join(
root, "{:05d}".format(scene_i), "level{}/optical_flow/*.npz".format(difficulty_level)
)))
if not len(flows):
flows = sorted(glob(osp.join(
root, "{:05d}".format(scene_i), "level{}/optical_flow/optical_flow/*.npz".format(difficulty_level)
)))
poses = osp.join(root, "{:05d}".format(scene_i), "level{}/pose.txt".format(difficulty_level))
poses_file = open(poses, 'r')
poses = poses_file.readlines()
frames = list(range(0, len(poses) - 1))
wrong_frames = osp.join(root, "{:05d}".format(scene_i), "level{}".format(difficulty_level),
"wrong_frames.txt")
if osp.exists(wrong_frames):
wrong_frames = open(wrong_frames, 'r')
wrong_frames = wrong_frames.read().splitlines()
wrong_frames = [i for i in wrong_frames if i != '']
wrong_frames = list(map(int, wrong_frames))
frames = (frame for frame in frames if frame not in wrong_frames)
for i in frames:
self.intrinsics_list += [intrinsics]
self.image_list += [[images[i], images[i + 1]]]
self.pose_list += [[poses[i], poses[i + 1]]]
self.depth_list += [[depths[i], depths[i + 1]]]
self.flow_list += [[flows[i], flows[i + 1]]]
def __len__(self):
return len(self.image_list)
def __getitem__(self, index):
if not self.init_seed:
worker_info = torch.utils.data.get_worker_info()
if worker_info is not None:
torch.manual_seed(worker_info.id)
np.random.seed(worker_info.id)
random.seed(worker_info.id)
self.init_seed = True
invert_order = np.random.rand() < self.invert_order_prob
# Read data
image1 = cv2.imread(self.image_list[index][0]) # BGR image
image2 = cv2.imread(self.image_list[index][1])
depth1 = cv2.imread(self.depth_list[index][0], cv2.IMREAD_ANYDEPTH)
depth2 = cv2.imread(self.depth_list[index][1], cv2.IMREAD_ANYDEPTH)
depth1 = np.array(depth1, dtype="float32") / (2 ** 16 - 1) * 30 # [m]
depth2 = np.array(depth2, dtype="float32") / (2 ** 16 - 1) * 30 # [m]
flow2d = np.load(self.flow_list[index][0])['optical_flow']
pose1 = self.pose_list[index][0].split()
pose2 = self.pose_list[index][1].split()
tx1, ty1, tz1 = float(pose1[1]), float(pose1[2]), float(pose1[3])
tx2, ty2, tz2 = float(pose2[1]), float(pose2[2]), float(pose2[3])
qx1, qy1, qz1, qw1 = float(pose1[4]), float(pose1[5]), float(pose1[6]), float(pose1[7])
qx2, qy2, qz2, qw2 = float(pose2[4]), float(pose2[5]), float(pose2[6]), float(pose2[7])
pose1 = np.array([tx1, ty1, tz1, qx1, qy1, qz1, qw1], dtype="float32") # Pose world-to-camera, openCV
pose2 = np.array([tx2, ty2, tz2, qx2, qy2, qz2, qw2], dtype="float32")
# From numpy to torch
image1 = torch.from_numpy(image1).float().permute(2, 0, 1) # RGB
image2 = torch.from_numpy(image2).float().permute(2, 0, 1)
depth1 = torch.from_numpy(depth1).float()
depth2 = torch.from_numpy(depth2).float()
flow2d = torch.from_numpy(flow2d).float()
pose1 = torch.from_numpy(pose1).float()
pose2 = torch.from_numpy(pose2).float()
intrinsics = torch.from_numpy(self.intrinsics_list[index]).float()
# Reduce resolution by res_factor
if self.res_factor != 1.:
h_i, w_i = depth1.shape
h, w = int(math.ceil(h_i / self.res_factor)), int(math.ceil(w_i / self.res_factor))
sy, sx = float(h) / float(h_i), float(w) / float(w_i)
image1 = F.interpolate(image1[None], [h, w], mode='bilinear', align_corners=True)[0]
image2 = F.interpolate(image2[None], [h, w], mode='bilinear', align_corners=True)[0]
depth1 = F.interpolate(depth1[None, None], [h, w], mode='bilinear', align_corners=True)[0, 0]
depth2 = F.interpolate(depth2[None, None], [h, w], mode='bilinear', align_corners=True)[0, 0]
flow2d = flow2d.permute(2, 0, 1)[None]
flow2d = F.interpolate(flow2d, [h, w], mode='bilinear', align_corners=True)[0]
flow2d = flow2d.permute(1, 2, 0) * torch.as_tensor([sx, sy])
intrinsics *= torch.as_tensor([sx, sy, sx, sy])
depth_mask = (depth1 > 0.1) * (depth1 < 30.0) * (depth2 > 0.1) * (depth2 < 30.0)
flow2d_mask = torch.sum(flow2d ** 2, dim=-1).sqrt() < 250
if self.depth_augmentor:
# Augment depth
s = 0.1 + 1.8 * np.random.rand()
pose1[:3] *= s
pose2[:3] *= s
depth1 = depth1 * s
depth2 = depth2 * s
else:
s = 1.0
if invert_order:
flowz = (1.0 / depth1 - 1.0 / depth2).unsqueeze(-1)
flow2d = - flowpy.forward_warp(flow2d.cpu().detach().numpy(), flow2d.cpu().detach().numpy())
flow2d = torch.from_numpy(flow2d).float()
flowxyz = torch.cat([flow2d, flowz], dim=-1) # flow2d + inverse depth change
else:
flowz = (1.0 / depth2 - 1.0 / depth1).unsqueeze(-1)
flowxyz = torch.cat([flow2d, flowz], dim=-1) # flow2d + inverse depth change
if self.augmentor:
image1, image2 = self.augmentor(image1, image2)
# Relative pose to go from pose1 to pose2
pose_gt = torch.from_numpy(pops.absolut_to_relative_poses(pose1, pose2)).float().to(pose1.device)
pose_gt_inv = pops.pose_from_matrix_to_quat(
torch.inverse(pops.pose_from_quat_to_matrix(pose_gt.unsqueeze(0)))).squeeze()
# Valid pixels mask
h, w = depth1.shape[:2]
y1, x1 = torch.meshgrid(
torch.arange(h).to(flow2d.device).float(),
torch.arange(w).to(flow2d.device).float())
x2 = x1 + flow2d[..., 0]
y2 = y1 + flow2d[..., 1]
coord2 = torch.stack([x2, y2], dim=-1)
_, valid_flow = bilinear_sampler(depth2.unsqueeze(0).unsqueeze(0), coord2.unsqueeze(0), mask=True)
if self.mode == 'train':
valid_mask = (depth_mask * flow2d_mask
* (flow2d[..., 0] > -120) * (flow2d[..., 0] < 120) * (flow2d[..., 1] > -120) * (
flow2d[..., 1] < 120)
* (valid_flow > 0).squeeze()).unsqueeze(-1)
elif self.mode == 'test':
valid_mask = depth_mask.unsqueeze(-1)
if not invert_order:
return image1, image2, depth1, depth2, pose_gt, intrinsics, flowxyz, valid_mask, s
else:
return image2, image1, depth2, depth1, pose_gt_inv, intrinsics, flowxyz, valid_mask, s