/
edgesDetectMexCNNFeats.cpp
256 lines (245 loc) · 8.72 KB
/
edgesDetectMexCNNFeats.cpp
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
/*******************************************************************************
* Structured Edge Detection Toolbox Version 3.0
* Copyright 2014 Piotr Dollar. [pdollar-at-microsoft.com]
* Please email me if you find bugs, or have suggestions or questions!
* Licensed under the MSR-LA Full Rights License [see license.txt]
*******************************************************************************/
#include <mex.h>
#include <math.h>
#include <stdlib.h>
#include <string.h>
#ifdef USEOMP
#include <omp.h>
#endif
typedef unsigned int uint32;
typedef unsigned short uint16;
typedef unsigned char uint8;
template<typename T> inline T min( T x, T y ) { return x<y ? x : y; }
// construct lookup array for mapping fids to channel indices
uint32* buildLookup( int *dims, int w )
{
int c, r, z, n=w*w*dims[2];
uint32 *cids=new uint32[n];
n=0;
for(z=0; z<dims[2]; z++)
for(c=0; c<w; c++)
for(r=0; r<w; r++)
cids[n++] = z*dims[0]*dims[1] + c*dims[0] + r;
return cids;
}
// [E,ind,segs] = mexFunction(model,I,chns,chnsSs) - helper for edgesDetect.m
void mexFunction( int nl, mxArray *pl[], int nr, const mxArray *pr[] )
{
// get inputs
mxArray *model = (mxArray*) pr[0];
float *I = (float*) mxGetData(pr[1]);
float *feats = (float*) mxGetData(pr[2]);
// extract relevant fields from model and options
float *thrs = (float*) mxGetData(mxGetField(model,0,"thrs"));
uint32 *fids = (uint32*) mxGetData(mxGetField(model,0,"fids"));
uint32 *child = (uint32*) mxGetData(mxGetField(model,0,"child"));
uint8 *segs = (uint8*) mxGetData(mxGetField(pr[0],0,"segs"));
uint8 *nSegs = (uint8*) mxGetData(mxGetField(pr[0],0,"nSegs"));
uint16 *eBins = (uint16*) mxGetData(mxGetField(model,0,"eBins"));
uint32 *eBnds = (uint32*) mxGetData(mxGetField(model,0,"eBnds"));
mxArray *opts = mxGetField(model,0,"opts");
const int shrink = (int) mxGetScalar(mxGetField(opts,0,"shrink"));
const int imWidth = (int) mxGetScalar(mxGetField(opts,0,"imWidth"));
const int gtWidth = (int) mxGetScalar(mxGetField(opts,0,"gtWidth"));
const int nChns = (int) mxGetScalar(mxGetField(opts,0,"nChns"));
const int nCells = (int) mxGetScalar(mxGetField(opts,0,"nCells"));
const uint32 nTotFtrs = (uint32) mxGetScalar(mxGetField(opts,0,"nTotFtrs"));
const int stride = (int) mxGetScalar(mxGetField(opts,0,"stride"));
const int nTreesEval = (int) mxGetScalar(mxGetField(opts,0,"nTreesEval"));
int sharpen = (int) mxGetScalar(mxGetField(opts,0,"sharpen"));
int nThreads = (int) mxGetScalar(mxGetField(opts,0,"nThreads"));
const int nBnds = int(mxGetNumberOfElements(mxGetField(model,0,"eBnds"))-1)/
int(mxGetNumberOfElements(mxGetField(model,0,"thrs")));
const char *msgSharpen="Model supports sharpening of at most %i pixels!\n";
if( sharpen>nBnds-1 )
{
sharpen=nBnds-1;
mexPrintf(msgSharpen,sharpen);
}
// get dimensions and constants
const mwSize *imgSize = mxGetDimensions(pr[1]);
const int h = (int) imgSize[0];
const int w = (int) imgSize[1];
const int Z = mxGetNumberOfDimensions(pr[1])<=2 ? 1 : imgSize[2];
const mwSize *fidsSize = mxGetDimensions(mxGetField(model,0,"fids"));
const int nTreeNodes = (int) fidsSize[0];
const int nTrees = (int) fidsSize[1];
const int h1 = (int) ceil(double(h-imWidth)/stride);
const int w1 = (int) ceil(double(w-imWidth)/stride);
const int h2 = h1*stride+gtWidth;
const int w2 = w1*stride+gtWidth;
const int imgDims[3] = {h,w,Z};
const int chnDims[3] = {h/shrink,w/shrink,nChns};
const int indDims[3] = {h1,w1,nTreesEval};
const int outDims[3] = {h2,w2,1};
const int segDims[5] = {gtWidth,gtWidth,h1,w1,nTreesEval};
// construct lookup tables
uint32 *iids, *eids, *cids;
iids = buildLookup( (int*)imgDims, gtWidth );
eids = buildLookup( (int*)outDims, gtWidth );
cids = buildLookup( (int*)chnDims, floor((double)2/shrink + 0.5));
//cids = buildLookup( (int*)chnDims, 1 );
// create outputs
pl[0] = mxCreateNumericArray(3,outDims,mxSINGLE_CLASS,mxREAL);
float *E = (float*) mxGetData(pl[0]);
pl[1] = mxCreateNumericArray(3,indDims,mxUINT32_CLASS,mxREAL);
uint32 *ind = (uint32*) mxGetData(pl[1]);
if(nl>2) pl[2] = mxCreateNumericArray(5,segDims,mxUINT8_CLASS,mxREAL);
uint8 *segsOut; if(nl>2) segsOut = (uint8*) mxGetData(pl[2]);
// apply forest to all patches and store leaf inds
#ifdef USEOMP
nThreads = min(nThreads,omp_get_max_threads());
#pragma omp parallel for num_threads(nThreads)
#endif
for( int c=0; c<w1; c++ )
for( int t=0; t<nTreesEval; t++ )
{
for( int r0=0; r0<2; r0++ )
for( int r=r0; r<h1; r+=2 )
{
int o = (r*stride/shrink) + (c*stride/shrink)*h/shrink;
// select tree to evaluate
int t1 = ((r+c)%2*nTreesEval+t)%nTrees;
uint32 k = t1*nTreeNodes;
while( child[k] )
{
// compute feature (either channel or self-similarity feature)
uint32 f = fids[k];
float ftr;
ftr = feats[cids[f]+o];
// compare ftr to threshold and move left or right accordingly
if( ftr < thrs[k] )
k = child[k]-1;
else
k = child[k];
k += t1*nTreeNodes;
}
// store leaf index and update edge maps
ind[ r + c*h1 + t*h1*w1 ] = k;
}
}
// compute edge maps (avoiding collisions from parallel executions)
if( !sharpen )
for( int c0=0; c0<gtWidth/stride; c0++ )
{
#ifdef USEOMP
#pragma omp parallel for num_threads(nThreads)
#endif
for( int c=c0; c<w1; c+=gtWidth/stride )
{
for( int r=0; r<h1; r++ )
for( int t=0; t<nTreesEval; t++ )
{
uint32 k = ind[ r + c*h1 + t*h1*w1 ];
float *E1 = E + (r*stride) + (c*stride)*h2;
int b0=eBnds[k*nBnds], b1=eBnds[k*nBnds+1];
if(b0==b1)
continue;
for( int b=b0; b<b1; b++ )
E1[eids[eBins[b]]]++;
if(nl>2)
memcpy(segsOut+(r+c*h1+t*h1*w1)*gtWidth*gtWidth,
segs+k*gtWidth*gtWidth,gtWidth*gtWidth*sizeof(uint8));
}
}
}
// computed sharpened edge maps, snapping to local color values
if( sharpen )
{
// compute neighbors array
const int g=gtWidth; uint16 N[4096*4];
for( int c=0; c<g; c++ )
for( int r=0; r<g; r++ )
{
int i=c*g+r;
uint16 *N1=N+i*4;
N1[0] = c>0 ? i-g : i; N1[1] = c<g-1 ? i+g : i;
N1[2] = r>0 ? i-1 : i; N1[3] = r<g-1 ? i+1 : i;
}
#ifdef USEOMP
#pragma omp parallel for num_threads(nThreads)
#endif
for( int c=0; c<w1; c++ )
for( int r=0; r<h1; r++ )
{
for( int t=0; t<nTreesEval; t++ )
{
// get current segment and copy into S
uint32 k = ind[ r + c*h1 + t*h1*w1 ];
int m = nSegs[k];
if( m==1 )
continue;
uint8 S0[4096], *S=(nl<=2) ? S0 : segsOut+(r+c*h1+t*h1*w1)*g*g;
memcpy(S,segs+k*g*g, g*g*sizeof(uint8));
// compute color model for each segment using every other pixel
int ci, ri, s, z;
float ns[100], mus[1000];
const float *I1 = I+(c*stride+(imWidth-g)/2)*h+r*stride+(imWidth-g)/2;
for( s=0; s<m; s++ )
{
ns[s]=0;
for( z=0; z<Z; z++ )
mus[s*Z+z]=0;
}
for( ci=0; ci<g; ci+=2 )
for( ri=0; ri<g; ri+=2 )
{
s = S[ci*g+ri];
ns[s]++;
for( z=0; z<Z; z++ )
mus[s*Z+z]+=I1[z*h*w+ci*h+ri];
}
for(s=0; s<m; s++)
for( z=0; z<Z; z++ )
mus[s*Z+z]/=ns[s];
// update segment S according to local color values
int b0=eBnds[k*nBnds], b1=eBnds[k*nBnds+sharpen];
for( int b=b0; b<b1; b++ )
{
float vs[10], d, e, eBest=1e10f;
int i, sBest=-1, ss[4];
for( i=0; i<4; i++ )
ss[i]=S[N[eBins[b]*4+i]];
for( z=0; z<Z; z++ )
vs[z]=I1[iids[eBins[b]]+z*h*w];
for( i=0; i<4; i++ )
{
s=ss[i];
if(s==sBest)
continue;
e=0;
for( z=0; z<Z; z++ )
{
d=mus[s*Z+z]-vs[z];
e+=d*d;
}
if( e<eBest )
{
eBest=e;
sBest=s;
}
}
S[eBins[b]]=sBest;
}
// convert mask to edge maps (examining expanded set of pixels)
float *E1 = E + c*stride*h2 + r*stride; b1=eBnds[k*nBnds+sharpen+1];
for( int b=b0; b<b1; b++ )
{
int i=eBins[b];
uint8 s=S[i];
uint16 *N1=N+i*4;
if( s!=S[N1[0]] || s!=S[N1[1]] || s!=S[N1[2]] || s!=S[N1[3]] )
E1[eids[i]]++;
}
}
}
}
// free memory
delete [] iids; delete [] eids;
delete [] cids;
}