-
Notifications
You must be signed in to change notification settings - Fork 71
/
utils.py
253 lines (215 loc) · 8.78 KB
/
utils.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
import matplotlib.patches as patches
from matplotlib.path import Path
import os
import sys
import io
import cv2
import time
import argparse
import shutil
import random
import zipfile
from glob import glob
import math
import numpy as np
import torch.nn.functional as F
import torchvision.transforms as transforms
from PIL import Image, ImageOps, ImageDraw, ImageFilter
import torch
import torchvision
import torch.nn as nn
import torch.distributed as dist
import matplotlib
from matplotlib import pyplot as plt
matplotlib.use('agg')
# #####################################################
# #####################################################
class ZipReader(object):
file_dict = dict()
def __init__(self):
super(ZipReader, self).__init__()
@staticmethod
def build_file_dict(path):
file_dict = ZipReader.file_dict
if path in file_dict:
return file_dict[path]
else:
file_handle = zipfile.ZipFile(path, 'r')
file_dict[path] = file_handle
return file_dict[path]
@staticmethod
def imread(path, image_name):
zfile = ZipReader.build_file_dict(path)
data = zfile.read(image_name)
im = Image.open(io.BytesIO(data))
return im
# ###########################################################################
# ###########################################################################
class GroupRandomHorizontalFlip(object):
"""Randomly horizontally flips the given PIL.Image with a probability of 0.5
"""
def __init__(self, is_flow=False):
self.is_flow = is_flow
def __call__(self, img_group, is_flow=False):
v = random.random()
if v < 0.5:
ret = [img.transpose(Image.FLIP_LEFT_RIGHT) for img in img_group]
if self.is_flow:
for i in range(0, len(ret), 2):
# invert flow pixel values when flipping
ret[i] = ImageOps.invert(ret[i])
return ret
else:
return img_group
class Stack(object):
def __init__(self, roll=False):
self.roll = roll
def __call__(self, img_group):
mode = img_group[0].mode
if mode == '1':
img_group = [img.convert('L') for img in img_group]
mode = 'L'
if mode == 'L':
return np.stack([np.expand_dims(x, 2) for x in img_group], axis=2)
elif mode == 'RGB':
if self.roll:
return np.stack([np.array(x)[:, :, ::-1] for x in img_group], axis=2)
else:
return np.stack(img_group, axis=2)
else:
raise NotImplementedError(f"Image mode {mode}")
class ToTorchFormatTensor(object):
""" Converts a PIL.Image (RGB) or numpy.ndarray (H x W x C) in the range [0, 255]
to a torch.FloatTensor of shape (C x H x W) in the range [0.0, 1.0] """
def __init__(self, div=True):
self.div = div
def __call__(self, pic):
if isinstance(pic, np.ndarray):
# numpy img: [L, C, H, W]
img = torch.from_numpy(pic).permute(2, 3, 0, 1).contiguous()
else:
# handle PIL Image
img = torch.ByteTensor(
torch.ByteStorage.from_buffer(pic.tobytes()))
img = img.view(pic.size[1], pic.size[0], len(pic.mode))
# put it from HWC to CHW format
# yikes, this transpose takes 80% of the loading time/CPU
img = img.transpose(0, 1).transpose(0, 2).contiguous()
img = img.float().div(255) if self.div else img.float()
return img
# ##########################################
# ##########################################
def create_random_shape_with_random_motion(video_length, imageHeight=240, imageWidth=432):
# get a random shape
height = random.randint(imageHeight//3, imageHeight-1)
width = random.randint(imageWidth//3, imageWidth-1)
edge_num = random.randint(6, 8)
ratio = random.randint(6, 8)/10
region = get_random_shape(
edge_num=edge_num, ratio=ratio, height=height, width=width)
region_width, region_height = region.size
# get random position
x, y = random.randint(
0, imageHeight-region_height), random.randint(0, imageWidth-region_width)
velocity = get_random_velocity(max_speed=3)
m = Image.fromarray(np.zeros((imageHeight, imageWidth)).astype(np.uint8))
m.paste(region, (y, x, y+region.size[0], x+region.size[1]))
masks = [m.convert('L')]
# return fixed masks
if random.uniform(0, 1) > 0.5:
return masks*video_length
# return moving masks
for _ in range(video_length-1):
x, y, velocity = random_move_control_points(
x, y, imageHeight, imageWidth, velocity, region.size, maxLineAcceleration=(3, 0.5), maxInitSpeed=3)
m = Image.fromarray(
np.zeros((imageHeight, imageWidth)).astype(np.uint8))
m.paste(region, (y, x, y+region.size[0], x+region.size[1]))
masks.append(m.convert('L'))
return masks
def get_random_shape(edge_num=9, ratio=0.7, width=432, height=240):
'''
There is the initial point and 3 points per cubic bezier curve.
Thus, the curve will only pass though n points, which will be the sharp edges.
The other 2 modify the shape of the bezier curve.
edge_num, Number of possibly sharp edges
points_num, number of points in the Path
ratio, (0, 1) magnitude of the perturbation from the unit circle,
'''
points_num = edge_num*3 + 1
angles = np.linspace(0, 2*np.pi, points_num)
codes = np.full(points_num, Path.CURVE4)
codes[0] = Path.MOVETO
# Using this instad of Path.CLOSEPOLY avoids an innecessary straight line
verts = np.stack((np.cos(angles), np.sin(angles))).T * \
(2*ratio*np.random.random(points_num)+1-ratio)[:, None]
verts[-1, :] = verts[0, :]
path = Path(verts, codes)
# draw paths into images
fig = plt.figure()
ax = fig.add_subplot(111)
patch = patches.PathPatch(path, facecolor='black', lw=2)
ax.add_patch(patch)
ax.set_xlim(np.min(verts)*1.1, np.max(verts)*1.1)
ax.set_ylim(np.min(verts)*1.1, np.max(verts)*1.1)
ax.axis('off') # removes the axis to leave only the shape
fig.canvas.draw()
# convert plt images into numpy images
data = np.frombuffer(fig.canvas.tostring_rgb(), dtype=np.uint8)
data = data.reshape((fig.canvas.get_width_height()[::-1] + (3,)))
plt.close(fig)
# postprocess
data = cv2.resize(data, (width, height))[:, :, 0]
data = (1 - np.array(data > 0).astype(np.uint8))*255
corrdinates = np.where(data > 0)
xmin, xmax, ymin, ymax = np.min(corrdinates[0]), np.max(
corrdinates[0]), np.min(corrdinates[1]), np.max(corrdinates[1])
region = Image.fromarray(data).crop((ymin, xmin, ymax, xmax))
return region
def random_accelerate(velocity, maxAcceleration, dist='uniform'):
speed, angle = velocity
d_speed, d_angle = maxAcceleration
if dist == 'uniform':
speed += np.random.uniform(-d_speed, d_speed)
angle += np.random.uniform(-d_angle, d_angle)
elif dist == 'guassian':
speed += np.random.normal(0, d_speed / 2)
angle += np.random.normal(0, d_angle / 2)
else:
raise NotImplementedError(
f'Distribution type {dist} is not supported.')
return (speed, angle)
def get_random_velocity(max_speed=3, dist='uniform'):
if dist == 'uniform':
speed = np.random.uniform(max_speed)
elif dist == 'guassian':
speed = np.abs(np.random.normal(0, max_speed / 2))
else:
raise NotImplementedError(
f'Distribution type {dist} is not supported.')
angle = np.random.uniform(0, 2 * np.pi)
return (speed, angle)
def random_move_control_points(X, Y, imageHeight, imageWidth, lineVelocity, region_size, maxLineAcceleration=(3, 0.5), maxInitSpeed=3):
region_width, region_height = region_size
speed, angle = lineVelocity
X += int(speed * np.cos(angle))
Y += int(speed * np.sin(angle))
lineVelocity = random_accelerate(
lineVelocity, maxLineAcceleration, dist='guassian')
if ((X > imageHeight - region_height) or (X < 0) or (Y > imageWidth - region_width) or (Y < 0)):
lineVelocity = get_random_velocity(maxInitSpeed, dist='guassian')
new_X = np.clip(X, 0, imageHeight - region_height)
new_Y = np.clip(Y, 0, imageWidth - region_width)
return new_X, new_Y, lineVelocity
# ##############################################
# ##############################################
if __name__ == '__main__':
trials = 10
for _ in range(trials):
video_length = 10
# The returned masks are either stationary (50%) or moving (50%)
masks = create_random_shape_with_random_motion(
video_length, imageHeight=240, imageWidth=432)
for m in masks:
cv2.imshow('mask', np.array(m))
cv2.waitKey(500)