Skip to content

Commit 19036e4

Browse files
art-programmerart-programmer
committed
minor changes
1 parent fb45b34 commit 19036e4

File tree

94 files changed

+4330
-3978841
lines changed

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

94 files changed

+4330
-3978841
lines changed

‎RecordReader.py

+83
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,83 @@
1+
import tensorflow as tf
2+
import numpy as np
3+
import threading
4+
import PIL.Image as Image
5+
from functools import partial
6+
from multiprocessing import Pool
7+
import cv2
8+
9+
import sys
10+
import os
11+
12+
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
13+
from modules import *
14+
15+
16+
HEIGHT=192
17+
WIDTH=256
18+
NUM_OBJECTS = 10
19+
NUM_THREADS = 4
20+
21+
22+
class RecordReader():
23+
def __init__(self):
24+
return
25+
26+
def getBatch(self, filename_queue, numOutputObjects = 20, batchSize = 16, min_after_dequeue = 1000, random=True, getLocal=False, getSegmentation=False, test=True):
27+
reader = tf.TFRecordReader()
28+
_, serialized_example = reader.read(filename_queue)
29+
30+
features = tf.parse_single_example(
31+
serialized_example,
32+
# Defaults are not specified since both keys are required.
33+
features={
34+
#'height': tf.FixedLenFeature([], tf.int64),
35+
#'width': tf.FixedLenFeature([], tf.int64),
36+
'image_raw': tf.FixedLenFeature([], tf.string),
37+
'image_path': tf.FixedLenFeature([], tf.string),
38+
'depth': tf.FixedLenFeature([HEIGHT * WIDTH], tf.float32),
39+
'info': tf.FixedLenFeature([4 * 4 + 4], tf.float32),
40+
'objects': tf.FixedLenFeature([NUM_OBJECTS * 13], tf.float32),
41+
})
42+
43+
# Convert from a scalar string tensor (whose single string has
44+
# length mnist.IMAGE_PIXELS) to a uint8 tensor with shape
45+
# [mnist.IMAGE_PIXELS].
46+
image = tf.decode_raw(features['image_raw'], tf.uint8)
47+
image = tf.cast(image, tf.float32) * (1. / 255) - 0.5
48+
image = tf.reshape(image, [HEIGHT, WIDTH, 3])
49+
50+
51+
depth = features['depth']
52+
depth = tf.reshape(depth, [HEIGHT, WIDTH, 1])
53+
54+
# normal = features['normal']
55+
# normal = tf.reshape(normal, [HEIGHT, WIDTH, 3])
56+
# normal = tf.nn.l2_normalize(normal, dim=2)
57+
58+
objects = tf.reshape(features['objects'], [NUM_OBJECTS, 13])
59+
numObjects = tf.reduce_sum(tf.cast(tf.greater(objects[:, 12], 0), dtype=tf.int32))
60+
61+
#numPlanes = tf.maximum(numPlanes, 1)
62+
#planes = tf.slice(planes, [0, 0], [numPlanes, 3])
63+
64+
if False:
65+
#shuffle_inds = tf.one_hot(tf.random_shuffle(tf.range(numPlanes)), depth = numPlanes)
66+
shuffle_inds = tf.one_hot(tf.range(numPlanes), numPlanes)
67+
68+
planes = tf.transpose(tf.matmul(tf.transpose(planes), shuffle_inds))
69+
planes = tf.reshape(planes, [numPlanes, 3])
70+
planes = tf.concat([planes, tf.zeros([numOutputObjects - numPlanes, 3])], axis=0)
71+
planes = tf.reshape(planes, [numOutputObjects, 3])
72+
pass
73+
74+
75+
76+
77+
if random:
78+
image_inp, object_inp, depth_gt, num_objects_gt, image_path, info = tf.train.shuffle_batch([image, objects, depth, numObjects, features['image_path'], features['info']], batch_size=batchSize, capacity=min_after_dequeue + (NUM_THREADS + 2) * batchSize, num_threads=NUM_THREADS, min_after_dequeue=min_after_dequeue)
79+
else:
80+
image_inp, object_inp, depth_gt, num_objects_gt, image_path, info = tf.train.batch([image, objects, depth, numObjects, features['image_path'], features['info']], batch_size=batchSize, capacity=min_after_dequeue + (NUM_THREADS + 2) * batchSize, num_threads=1)
81+
pass
82+
global_gt_dict = {'object': object_inp, 'depth': depth_gt, 'num_objects': num_objects_gt, 'image_path': image_path, 'info': info}
83+
return image_inp, global_gt_dict, {}

‎RecordReaderLocal.py

+154
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,154 @@
1+
import tensorflow as tf
2+
import numpy as np
3+
import threading
4+
import PIL.Image as Image
5+
from functools import partial
6+
from multiprocessing import Pool
7+
import cv2
8+
9+
import sys
10+
import os
11+
12+
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
13+
from modules import *
14+
15+
16+
HEIGHT=192
17+
WIDTH=256
18+
NUM_OBJECTS = 10
19+
NUM_THREADS = 4
20+
21+
22+
class RecordReader():
23+
def __init__(self):
24+
return
25+
26+
def getBatch(self, filename_queue, options, min_after_dequeue = 1000, random=True):
27+
reader = tf.TFRecordReader()
28+
_, serialized_example = reader.read(filename_queue)
29+
30+
features = tf.parse_single_example(
31+
serialized_example,
32+
# Defaults are not specified since both keys are required.
33+
features={
34+
#'height': tf.FixedLenFeature([], tf.int64),
35+
#'width': tf.FixedLenFeature([], tf.int64),
36+
'image_raw': tf.FixedLenFeature([], tf.string),
37+
'image_path': tf.FixedLenFeature([], tf.string),
38+
'depth': tf.FixedLenFeature([HEIGHT * WIDTH], tf.float32),
39+
'info': tf.FixedLenFeature([4 * 4 + 4], tf.float32),
40+
'objects': tf.FixedLenFeature([NUM_OBJECTS * 13], tf.float32),
41+
})
42+
43+
# Convert from a scalar string tensor (whose single string has
44+
# length mnist.IMAGE_PIXELS) to a uint8 tensor with shape
45+
# [mnist.IMAGE_PIXELS].
46+
image = tf.decode_raw(features['image_raw'], tf.uint8)
47+
image = tf.cast(image, tf.float32) * (1. / 255) - 0.5
48+
image = tf.reshape(image, [HEIGHT, WIDTH, 3])
49+
50+
51+
depth = features['depth']
52+
depth = tf.reshape(depth, [HEIGHT, WIDTH, 1])
53+
54+
# normal = features['normal']
55+
# normal = tf.reshape(normal, [HEIGHT, WIDTH, 3])
56+
# normal = tf.nn.l2_normalize(normal, dim=2)
57+
58+
objects = tf.reshape(features['objects'], [NUM_OBJECTS, 13])
59+
60+
outputWidth = WIDTH / options.outputStride
61+
outputHeight = HEIGHT / options.outputStride
62+
63+
if options.axisAligned or options.useAnchor:
64+
centers = objects[:, :3]
65+
sizes = objects[:, 3:6]
66+
towards = objects[:, 6:9]
67+
up = objects[:, 9:12]
68+
right = tf.cross(towards, up)
69+
right /= tf.maximum(tf.norm(right, axis=-1, keep_dims=True), 1e-4)
70+
directionsArray = np.array([[1, 1, 1],
71+
[1, 1, -1],
72+
[1, -1, 1],
73+
[1, -1, -1],
74+
[-1, 1, 1],
75+
[-1, 1, -1],
76+
[-1, -1, 1],
77+
[-1, -1, -1]])
78+
directions = tf.expand_dims(tf.constant(directionsArray.reshape(-1), shape=directionsArray.shape), 0)
79+
cornerPoints = tf.expand_dims(centers, 1) + tf.matmul(tf.expand_dims(sizes, 1) / 2 * directionsArray, tf.stack([towards, up, right], axis=1))
80+
maxs = tf.reduce_max(cornerPoints, axis=1)
81+
mins = tf.reduce_min(cornerPoints, axis=1)
82+
sizes = maxs - mins
83+
#centers = (mins + maxs) / 2
84+
objects = tf.concat([centers, sizes, objects[:, 6:]], axis=1)
85+
pass
86+
87+
88+
numObjects = tf.reduce_sum(tf.cast(tf.greater(objects[:, 12], 0), dtype=tf.int32))
89+
info = features['info']
90+
91+
objects = objects[:numObjects]
92+
93+
if options.useAnchor:
94+
centers = objects[:, :3] - objects[:, 3:6] / 2 * tf.expand_dims(tf.constant([1, 0, 0], dtype=tf.float32), 0)
95+
else:
96+
centers = objects[:, :3]
97+
pass
98+
99+
U = (centers[:, 2] / centers[:, 0] * info[0] + info[2]) / info[16]
100+
V = tf.clip_by_value((-centers[:, 1] / centers[:, 0] * info[5] + info[6]) / info[17], 0, 1)
101+
102+
if True:
103+
validMask = tf.logical_and(tf.logical_and(tf.logical_and(tf.greater_equal(U, 0), tf.less_equal(U, 1)), tf.logical_and(tf.greater_equal(V, 0), tf.less_equal(V, 1))), tf.greater(centers[:, 0], 0))
104+
objects = tf.boolean_mask(objects, validMask)
105+
U = tf.boolean_mask(U, validMask)
106+
V = tf.boolean_mask(V, validMask)
107+
numObjects = tf.reduce_sum(tf.cast(validMask, tf.int32))
108+
pass
109+
110+
U = tf.clip_by_value(U, 0, 1)
111+
V = tf.clip_by_value(V, 0, 1)
112+
113+
gridU = tf.clip_by_value(tf.cast((U * outputWidth), tf.int32), 0, outputWidth - 1)
114+
gridV = tf.clip_by_value(tf.cast((V * outputHeight), tf.int32), 0, outputHeight - 1)
115+
indices = gridV * outputWidth + gridU
116+
local_scores = tf.reshape(tf.maximum(tf.unsorted_segment_max(tf.ones([numObjects]), indices, num_segments=outputWidth * outputHeight), 0), (outputHeight, outputWidth, 1))
117+
class_prob = tf.one_hot(tf.cast(objects[:, 12], tf.int32), depth=40, axis=-1)
118+
local_classes = tf.reshape(tf.maximum(tf.unsorted_segment_max(class_prob, indices, num_segments=outputWidth * outputHeight), 0), (outputHeight, outputWidth, -1))
119+
120+
if options.useAnchor:
121+
mins = objects[:, :3] - objects[:, 3:6] / 2 * tf.expand_dims(tf.constant([1, -1, 1], dtype=tf.float32), 0)
122+
maxs = objects[:, :3] - objects[:, 3:6] / 2 * tf.expand_dims(tf.constant([1, 1, -1], dtype=tf.float32), 0)
123+
minU = (mins[:, 2] / mins[:, 0] * info[0] + info[2]) / info[16]
124+
minV = (-mins[:, 1] / mins[:, 0] * info[5] + info[6]) / info[17]
125+
maxU = (maxs[:, 2] / maxs[:, 0] * info[0] + info[2]) / info[16]
126+
maxV = (-maxs[:, 1] / maxs[:, 0] * info[5] + info[6]) / info[17]
127+
boxes = tf.stack([U, V, maxU - minU, maxV - minV], axis=-1)
128+
boxes = tf.reshape(tf.maximum(tf.unsorted_segment_max(boxes, indices, num_segments=outputWidth * outputHeight), 0), (outputHeight, outputWidth, 4))
129+
130+
anchorW = tf.fill((outputHeight, outputWidth, 1), 1.0 / outputWidth)
131+
anchorH = tf.fill((outputHeight, outputWidth, 1), 1.0 / outputHeight)
132+
anchors = tf.stack([tf.tile(tf.expand_dims(tf.range(outputWidth, dtype=tf.float32), 0), (outputHeight, 1)) / outputWidth, tf.tile(tf.expand_dims(tf.range(outputHeight, dtype=tf.float32), 1), (1, outputWidth)) / outputHeight], axis=-1)
133+
anchors = tf.concat([anchors, anchorW, anchorH], axis=-1)
134+
135+
#local_parameters = tf.reshape(tf.unsorted_segment_sum(objects[:, :12], indices, num_segments=outputWidth * outputHeight), (outputHeight, outputWidth, 12))
136+
depths = tf.clip_by_value(tf.reshape(tf.maximum(tf.unsorted_segment_max(tf.stack([objects[:, 0] - objects[:, 3] / 2, objects[:, 0] + objects[:, 3] / 2], axis=-1), indices, num_segments=outputWidth * outputHeight), 0), (outputHeight, outputWidth, 2)), 0, 10)
137+
local_parameters = tf.concat([(boxes[:, :, :2] - anchors[:, :, :2]) / anchors[:, :, 2:4], tf.minimum(boxes[:, :, 2:4] / anchors[:, :, 2:4], options.outputStride * 2), depths], axis=-1)
138+
local_parameters *= local_scores
139+
local_parameters = tf.concat([local_parameters, tf.zeros(local_parameters.shape)], axis=-1)
140+
#local_parameters = tf.reshape(tf.unsorted_segment_sum(objects[:, :12], indices, num_segments=outputWidth * outputHeight), (outputHeight, outputWidth, 12))
141+
else:
142+
local_parameters = tf.reshape(tf.unsorted_segment_sum(objects[:, :12], indices, num_segments=outputWidth * outputHeight), (outputHeight, outputWidth, 12))
143+
pass
144+
145+
objects = tf.reshape(tf.concat([objects, tf.zeros((NUM_OBJECTS - numObjects, 13))], axis=0), (NUM_OBJECTS, 13))
146+
147+
if random:
148+
image_inp, object_inp, depth_gt, num_objects_gt, image_path, info, local_scores_gt, local_parameters_gt, local_classes_gt = tf.train.shuffle_batch([image, objects, depth, numObjects, features['image_path'], features['info'], local_scores, local_parameters, local_classes], batch_size=options.batchSize, capacity=min_after_dequeue + (NUM_THREADS + 2) * options.batchSize, num_threads=NUM_THREADS, min_after_dequeue=min_after_dequeue)
149+
else:
150+
image_inp, object_inp, depth_gt, num_objects_gt, image_path, info, local_scores_gt, local_parameters_gt, local_classes_gt = tf.train.batch([image, objects, depth, numObjects, features['image_path'], features['info'], local_scores, local_parameters, local_classes], batch_size=options.batchSize, capacity=min_after_dequeue + (NUM_THREADS + 2) * options.batchSize, num_threads=1)
151+
pass
152+
global_gt_dict = {'object': object_inp, 'depth': depth_gt, 'num_objects': num_objects_gt}
153+
local_gt_dict = {'score': local_scores_gt, 'object': local_parameters_gt, 'image_path': image_path, 'info': info, 'num_objects': num_objects_gt, 'class': local_classes_gt}
154+
return image_inp, global_gt_dict, local_gt_dict

0 commit comments

Comments
 (0)