-
Notifications
You must be signed in to change notification settings - Fork 116
/
HogExtractor.scala
296 lines (262 loc) · 10.3 KB
/
HogExtractor.scala
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
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
/**
* -------------------------------------------------------
* Copyright (C) 2014 Henry Milner
* Copyright (C) 2011-2012 Ross Girshick
* Copyright (C) 2008, 2009, 2010 Pedro Felzenszwalb, Ross Girshick
* Copyright (C) 2007 Pedro Felzenszwalb, Deva Ramanan
*
* This file is part of the voc-releaseX code
* (http://people.cs.uchicago.edu/~rbg/latent/)
* and is available under the terms of an MIT-like license
* provided in COPYING. Please retain this notice and
* COPYING if you use this file (or a portion of it) in
* your project.
* -------------------------------------------------------
*/
package keystoneml.nodes.images
import breeze.linalg._
import keystoneml.workflow.Transformer
import keystoneml.utils.ChannelMajorArrayVectorizedImage
import keystoneml.utils.Image
import keystoneml.utils.ImageUtils
/**
* Histogram of Gradients (HoG)
*
* Translated by Henry Milner from C code originally written by Ross Girshick.
* The original code is available at:
* https://github.com/rbgirshick/voc-dpm/blob/master/features/features.cc
*/
class HogExtractor(binSize: Int) extends Transformer[Image, DenseMatrix[Float]] {
// small value, used to avoid division by zero
private val EPSILON = 0.0001
// unit vectors used to compute gradient orientation
private val uu = Array(
1.0000,
0.9397,
0.7660,
0.500,
0.1736,
-0.1736,
-0.5000,
-0.7660,
-0.9397)
private val vv = Array(
0.0000,
0.3420,
0.6428,
0.8660,
0.9848,
0.9848,
0.8660,
0.6428,
0.3420)
private val numOrientations = uu.length
def apply(image: Image): DenseMatrix[Float] = {
val numXCells = math.round(image.metadata.xDim.toDouble / binSize).toInt
val numYCells = math.round(image.metadata.yDim.toDouble / binSize).toInt
val numChannels = image.metadata.numChannels
val hist = computeHist(image, numXCells, numYCells, numChannels, binSize)
val norm = computeNormsFromHist(hist, numXCells, numYCells)
computeFeaturesFromHist(hist, norm, numXCells, numYCells)
}
def computeHist(
image: Image,
numXCells: Int,
numYCells: Int,
numChannels: Int,
binSize: Int): Array[Float] = {
// A flat array conceptually indexed by an (x, y, orientationIdx) triple. To
// reference the element at (x, y, o) use hist(x + y*numXCells + o*numYCells*numXCells).
val hist: Array[Float] = Array.ofDim(numXCells * numYCells * 18)
val numVisibleXPixels = numXCells * binSize
val numVisibleYPixels = numYCells * binSize
var x = 1
while (x < numVisibleXPixels - 1) {
var y = 1
while (y < numVisibleYPixels - 1) {
// First we compute the channel with the highest-magnitude gradient,
// and throw away the other channels for this pixel.
var highestMagnitudeChannel = -1
var bestChannelMagnitudeSquared = Double.NegativeInfinity
var bestChannelDx = Double.NegativeInfinity
var bestChannelDy = Double.NegativeInfinity
var channelIdx = 2
while (channelIdx >= 0) {
val dx = image.get(x + 1, y, channelIdx) - image.get(x - 1, y, channelIdx)
val dy = image.get(x, y + 1, channelIdx) - image.get(x, y - 1, channelIdx)
val magnitudeSquared = dx*dx + dy*dy
if (magnitudeSquared > bestChannelMagnitudeSquared) {
highestMagnitudeChannel = channelIdx
bestChannelMagnitudeSquared = magnitudeSquared
bestChannelDx = dx
bestChannelDy = dy
}
channelIdx -= 1
}
val dx = bestChannelDx
val dy = bestChannelDy
val magnitude = math.sqrt(bestChannelMagnitudeSquared)
// We snap to one of 18 orientations.
var bestOrientationDot = 0.0 //Double.NegativeInfinity
var bestOrientationIdx = 0
var orientationIdx = 0
while (orientationIdx < numOrientations) {
val dot = uu(orientationIdx)*dy + vv(orientationIdx)*dx
if (dot > bestOrientationDot) {
bestOrientationIdx = orientationIdx
bestOrientationDot = dot
} else if (-dot > bestOrientationDot) {
bestOrientationIdx = orientationIdx + numOrientations
bestOrientationDot = -dot
}
orientationIdx += 1
}
// We add to 4 histograms around the pixel (x,y) using bilinear interpolation
// TODO: Clean this. It's just rounding some stuff to the nearest bin,
// but it looks way more mysterious than it should.
val yp = (y + 0.5)/binSize - 0.5
val xp = (x + 0.5)/binSize - 0.5
val iyp = math.floor(yp).toInt
val ixp = math.floor(xp).toInt
val vy0 = yp - iyp
val vx0 = xp - ixp
val vy1 = 1.0 - vy0
val vx1 = 1.0 - vx0
if (iyp >= 0 && ixp >= 0) {
hist(ixp + iyp*numXCells + bestOrientationIdx*numXCells*numYCells) +=
(vy1*vx1*magnitude).toFloat
}
if (iyp + 1 < numYCells && ixp >= 0) {
hist(ixp + (iyp + 1)*numXCells + bestOrientationIdx*numXCells*numYCells) +=
(vy0*vx1*magnitude).toFloat
}
if (iyp >= 0 && ixp + 1 < numXCells) {
hist((ixp + 1) + iyp*numXCells + bestOrientationIdx*numXCells*numYCells) +=
(vy1*vx0*magnitude).toFloat
}
if (iyp + 1 < numYCells && ixp + 1 < numXCells) {
hist((ixp + 1) + (iyp + 1)*numXCells + bestOrientationIdx*numXCells*numYCells) +=
(vy0*vx0*magnitude).toFloat
}
y += 1
}
x += 1
}
hist
}
def computeNormsFromHist(hist: Array[Float], numXCells: Int, numYCells: Int): Array[Float] = {
// A flat array conceptually indexed by an (x, y) pair. To reference
// the element at (x, y) use hist(x + y*numXCells).
val norm: Array[Float] = Array.ofDim(numXCells * numYCells)
// Compute energy in each block by summing over orientations.
// Values for opposite orientations are combined.
// norm(x, y) = sum_{o < 9} (hist(x, y, o) + hist(x, y, o+9))^2
var o = 0
while (o < 9) {
val oppositeO = o + 9
var y = 0
while (y < numYCells) {
var x = 0
while (x < numXCells) {
val histValueO = hist(x + y*numXCells + o*numXCells*numYCells)
val histValueOppositeO = hist(x + y*numXCells + oppositeO*numXCells*numYCells)
norm(x + y*numXCells) +=
(histValueO + histValueOppositeO) * (histValueO + histValueOppositeO)
x += 1
}
y += 1
}
o += 1
}
norm
}
private def computeFeaturesFromHist(
hist: Array[Float],
norm: Array[Float],
numXCells: Int,
numYCells: Int): DenseMatrix[Float] = {
val numXCellsWithFeatures = math.max(numXCells-2, 0)
val numYCellsWithFeatures = math.max(numYCells-2, 0)
val numFeatures = 27 + 4 + 1
// Indexed by (x + y*numXCellsWithFeatures, featureIdx)
val features = new DenseMatrix[Float](
numXCellsWithFeatures * numYCellsWithFeatures, numFeatures)
var x = 0
while (x < numXCellsWithFeatures) {
var y = 0
while (y < numYCellsWithFeatures) {
val featureDestinationStartingIdx = y + x*numYCellsWithFeatures
val n1NormOffset = (y + 1)*numXCells + (x + 1)
val n1 = 1.0 / math.sqrt(norm(n1NormOffset) + norm(n1NormOffset + 1) +
norm(n1NormOffset + numXCells) + norm(n1NormOffset + numXCells + 1) + EPSILON)
val n2NormOffset = (y + 1)*numXCells + x
val n2 = 1.0 / math.sqrt(norm(n2NormOffset) + norm(n2NormOffset + 1) +
norm(n2NormOffset + numXCells) + norm(n2NormOffset + numXCells + 1) + EPSILON)
val n3NormOffset = y*numXCells + (x + 1)
val n3 = 1.0 / math.sqrt(norm(n3NormOffset) + norm(n3NormOffset + 1) +
norm(n3NormOffset + numXCells) + norm(n3NormOffset + numXCells + 1) + EPSILON)
val n4NormOffset = y*numXCells + x
val n4 = 1.0 / math.sqrt(norm(n4NormOffset) + norm(n4NormOffset + 1) +
norm(n4NormOffset + numXCells) + norm(n4NormOffset + numXCells + 1) + EPSILON)
var t1 = 0.0
var t2 = 0.0
var t3 = 0.0
var t4 = 0.0
// Contrast-sensitive features
// TODO: Clean up the indexing here; it uses pointer-incrementing style
// rather than calculating indices explicitly, which would be less risky,
// clearer, and probably not significantly less performant.
var histOffset0 = (y + 1)*numXCells + (x + 1)
var o0 = 0
var featureOffset0 = 0
while (o0 < 18) {
val h1 = math.min(hist(histOffset0) * n1, 0.2)
val h2 = math.min(hist(histOffset0) * n2, 0.2)
val h3 = math.min(hist(histOffset0) * n3, 0.2)
val h4 = math.min(hist(histOffset0) * n4, 0.2)
features(featureDestinationStartingIdx, featureOffset0) =
(0.5 * (h1 + h2 + h3 + h4)).toFloat
t1 += h1
t2 += h2
t3 += h3
t4 += h4
featureOffset0 += 1
histOffset0 += (numXCells * numYCells)
o0 += 1
}
// contrast-insensitive features
var o1 = 0
var histOffset1 = (y + 1)*numXCells + (x + 1)
var featureOffset1 = featureOffset0
while (o1 < 9) {
val sum = hist(histOffset1) + hist(histOffset1 + 9*numXCells*numYCells)
val h1 = math.min(sum * n1, 0.2)
val h2 = math.min(sum * n2, 0.2)
val h3 = math.min(sum * n3, 0.2)
val h4 = math.min(sum * n4, 0.2)
features(featureDestinationStartingIdx, featureOffset1) =
(0.5 * (h1 + h2 + h3 + h4)).toFloat
featureOffset1 += 1
histOffset1 += (numXCells * numYCells)
o1 += 1
}
// texture features
var featureOffset2 = featureOffset1
features(featureDestinationStartingIdx, featureOffset2) = (0.2357 * t1).toFloat
featureOffset2 += 1
features(featureDestinationStartingIdx, featureOffset2) = (0.2357 * t2).toFloat
featureOffset2 += 1
features(featureDestinationStartingIdx, featureOffset2) = (0.2357 * t3).toFloat
featureOffset2 += 1
features(featureDestinationStartingIdx, featureOffset2) = (0.2357 * t4).toFloat
// truncation feature
var featureOffset3 = featureOffset2
featureOffset3 += 1
features(featureDestinationStartingIdx, featureOffset3) = 0
y += 1
}
x += 1
}
features
}
}