root / rgbdslam / external / siftgpu / src / SiftGPU / SiftMatch.cpp @ 9240aaa3
History | View | Annotate | Download (20.8 KB)
1 |
////////////////////////////////////////////////////////////////////////////
|
---|---|
2 |
// File: SiftMatch.cpp
|
3 |
// Author: Changchang Wu
|
4 |
// Description : implementation of SiftMatchGPU and SiftMatchGL
|
5 |
//
|
6 |
//
|
7 |
// Copyright (c) 2007 University of North Carolina at Chapel Hill
|
8 |
// All Rights Reserved
|
9 |
//
|
10 |
// Permission to use, copy, modify and distribute this software and its
|
11 |
// documentation for educational, research and non-profit purposes, without
|
12 |
// fee, and without a written agreement is hereby granted, provided that the
|
13 |
// above copyright notice and the following paragraph appear in all copies.
|
14 |
//
|
15 |
// The University of North Carolina at Chapel Hill make no representations
|
16 |
// about the suitability of this software for any purpose. It is provided
|
17 |
// 'as is' without express or implied warranty.
|
18 |
//
|
19 |
// Please send BUG REPORTS to ccwu@cs.unc.edu
|
20 |
//
|
21 |
////////////////////////////////////////////////////////////////////////////
|
22 |
|
23 |
|
24 |
#include "GL/glew.h" |
25 |
#include <iostream> |
26 |
#include <iomanip> |
27 |
#include <vector> |
28 |
#include <strstream> |
29 |
#include <algorithm> |
30 |
using namespace std; |
31 |
#include <string.h> |
32 |
#include "GlobalUtil.h" |
33 |
|
34 |
#include "ProgramGLSL.h" |
35 |
#include "GLTexImage.h" |
36 |
#include "SiftGPU.h" |
37 |
#include "SiftMatch.h" |
38 |
#include "FrameBufferObject.h" |
39 |
|
40 |
#if defined(CUDA_SIFTGPU_ENABLED)
|
41 |
#include "CuTexImage.h" |
42 |
#include "SiftMatchCU.h" |
43 |
#endif
|
44 |
|
45 |
|
46 |
SiftMatchGL::SiftMatchGL(int max_sift, int use_glsl): SiftMatchGPU() |
47 |
{ |
48 |
s_multiply = s_col_max = s_row_max = s_guided_mult = NULL;
|
49 |
_num_sift[0] = _num_sift[1] = 0; |
50 |
_id_sift[0] = _id_sift[1] = 0; |
51 |
_have_loc[0] = _have_loc[1] = 0; |
52 |
_max_sift = max_sift <=0 ? 4096 : ((max_sift + 31)/ 32 * 32) ; |
53 |
_pixel_per_sift = 32; //must be 32 |
54 |
_sift_num_stripe = 1;
|
55 |
_sift_per_stripe = 1;
|
56 |
_sift_per_row = _sift_per_stripe * _sift_num_stripe; |
57 |
_initialized = 0;
|
58 |
} |
59 |
|
60 |
SiftMatchGL::~SiftMatchGL() |
61 |
{ |
62 |
if(s_multiply) delete s_multiply; |
63 |
if(s_guided_mult) delete s_guided_mult; |
64 |
if(s_col_max) delete s_col_max; |
65 |
if(s_row_max) delete s_row_max; |
66 |
} |
67 |
|
68 |
void SiftMatchGL::SetMaxSift(int max_sift) |
69 |
{ |
70 |
|
71 |
max_sift = ((max_sift + 31)/32)*32; |
72 |
if(max_sift > GlobalUtil::_texMaxDimGL) max_sift = GlobalUtil::_texMaxDimGL;
|
73 |
if(max_sift > _max_sift)
|
74 |
{ |
75 |
_max_sift = max_sift; |
76 |
AllocateSiftMatch(); |
77 |
_have_loc[0] = _have_loc[1] = 0; |
78 |
_id_sift[0] = _id_sift[1] = -1; |
79 |
_num_sift[0] = _num_sift[1] = 1; |
80 |
}else
|
81 |
{ |
82 |
_max_sift = max_sift; |
83 |
} |
84 |
|
85 |
} |
86 |
|
87 |
void SiftMatchGL::AllocateSiftMatch()
|
88 |
{ |
89 |
//parameters, number of sift is limited by the texture size
|
90 |
if(_max_sift > GlobalUtil::_texMaxDimGL) _max_sift = GlobalUtil::_texMaxDimGL;
|
91 |
///
|
92 |
int h = _max_sift / _sift_per_row;
|
93 |
int n = (GlobalUtil::_texMaxDimGL + h - 1) / GlobalUtil::_texMaxDimGL; |
94 |
if ( n > 1) {_sift_num_stripe *= n; _sift_per_row *= n; } |
95 |
|
96 |
//initialize
|
97 |
|
98 |
_texDes[0].InitTexture(_sift_per_row * _pixel_per_sift, _max_sift / _sift_per_row, 0,GL_RGBA8); |
99 |
_texDes[1].InitTexture(_sift_per_row * _pixel_per_sift, _max_sift / _sift_per_row, 0, GL_RGBA8); |
100 |
_texLoc[0].InitTexture(_sift_per_row , _max_sift / _sift_per_row, 0); |
101 |
_texLoc[1].InitTexture(_sift_per_row , _max_sift / _sift_per_row, 0); |
102 |
|
103 |
if(GlobalUtil::_SupportNVFloat || GlobalUtil::_SupportTextureRG)
|
104 |
{ |
105 |
//use single-component texture to save memory
|
106 |
#ifndef GL_R32F
|
107 |
#define GL_R32F 0x822E |
108 |
#endif
|
109 |
GLuint format = GlobalUtil::_SupportNVFloat ? GL_FLOAT_R_NV : GL_R32F; |
110 |
_texDot.InitTexture(_max_sift, _max_sift, 0, format);
|
111 |
_texMatch[0].InitTexture(16, _max_sift / 16, 0, format); |
112 |
_texMatch[1].InitTexture(16, _max_sift / 16, 0, format); |
113 |
}else
|
114 |
{ |
115 |
_texDot.InitTexture(_max_sift, _max_sift, 0);
|
116 |
_texMatch[0].InitTexture(16, _max_sift / 16, 0); |
117 |
_texMatch[1].InitTexture(16, _max_sift / 16, 0); |
118 |
} |
119 |
|
120 |
} |
121 |
void SiftMatchGL::InitSiftMatch()
|
122 |
{ |
123 |
if(_initialized) return; |
124 |
GlobalUtil::InitGLParam(0);
|
125 |
if(GlobalUtil::_GoodOpenGL == 0) return; |
126 |
AllocateSiftMatch(); |
127 |
LoadSiftMatchShadersGLSL(); |
128 |
_initialized = 1;
|
129 |
} |
130 |
|
131 |
|
132 |
void SiftMatchGL::SetDescriptors(int index, int num, const unsigned char* descriptors, int id) |
133 |
{ |
134 |
if(_initialized == 0) return; |
135 |
if (index > 1) index = 1; |
136 |
if (index < 0) index = 0; |
137 |
_have_loc[index] = 0;
|
138 |
|
139 |
//the same feature is already set
|
140 |
if(id !=-1 && id == _id_sift[index]) return ; |
141 |
_id_sift[index] = id; |
142 |
|
143 |
if(num > _max_sift) num = _max_sift;
|
144 |
|
145 |
sift_buffer.resize(num * 128 /4); |
146 |
memcpy(&sift_buffer[0], descriptors, 128 * num); |
147 |
_num_sift[index] = num; |
148 |
int w = _sift_per_row * _pixel_per_sift;
|
149 |
int h = (num + _sift_per_row - 1)/ _sift_per_row; |
150 |
sift_buffer.resize(w * h * 4, 0); |
151 |
_texDes[index].SetImageSize(w , h); |
152 |
_texDes[index].BindTex(); |
153 |
if(_sift_num_stripe == 1) |
154 |
{ |
155 |
glTexSubImage2D(GlobalUtil::_texTarget, 0, 0, 0, w, h, GL_RGBA, GL_UNSIGNED_BYTE, &sift_buffer[0]); |
156 |
}else
|
157 |
{ |
158 |
for(int i = 0; i < _sift_num_stripe; ++i) |
159 |
{ |
160 |
int ws = _sift_per_stripe * _pixel_per_sift;
|
161 |
int x = i * ws;
|
162 |
int pos = i * ws * h * 4; |
163 |
glTexSubImage2D(GlobalUtil::_texTarget, 0, x, 0, ws, h, GL_RGBA, GL_UNSIGNED_BYTE, &sift_buffer[pos]); |
164 |
} |
165 |
} |
166 |
_texDes[index].UnbindTex(); |
167 |
|
168 |
} |
169 |
|
170 |
void SiftMatchGL::SetFeautreLocation(int index, const float* locations, int gap) |
171 |
{ |
172 |
if(_num_sift[index] <=0) return; |
173 |
int w = _sift_per_row ;
|
174 |
int h = (_num_sift[index] + _sift_per_row - 1)/ _sift_per_row; |
175 |
sift_buffer.resize(_num_sift[index] * 2);
|
176 |
if(gap == 0) |
177 |
{ |
178 |
memcpy(&sift_buffer[0], locations, _num_sift[index] * 2 * sizeof(float)); |
179 |
}else
|
180 |
{ |
181 |
for(int i = 0; i < _num_sift[index]; ++i) |
182 |
{ |
183 |
sift_buffer[i*2] = *locations++;
|
184 |
sift_buffer[i*2+1]= *locations ++; |
185 |
locations += gap; |
186 |
} |
187 |
} |
188 |
sift_buffer.resize(w * h * 2, 0); |
189 |
_texLoc[index].SetImageSize(w , h); |
190 |
_texLoc[index].BindTex(); |
191 |
if(_sift_num_stripe == 1) |
192 |
{ |
193 |
glTexSubImage2D(GlobalUtil::_texTarget, 0, 0, 0, w, h, GL_LUMINANCE_ALPHA , GL_FLOAT , &sift_buffer[0]); |
194 |
}else
|
195 |
{ |
196 |
for(int i = 0; i < _sift_num_stripe; ++i) |
197 |
{ |
198 |
int ws = _sift_per_stripe;
|
199 |
int x = i * ws;
|
200 |
int pos = i * ws * h * 2; |
201 |
glTexSubImage2D(GlobalUtil::_texTarget, 0, x, 0, ws, h, GL_LUMINANCE_ALPHA , GL_FLOAT, &sift_buffer[pos]); |
202 |
} |
203 |
} |
204 |
_texLoc[index].UnbindTex(); |
205 |
_have_loc[index] = 1;
|
206 |
} |
207 |
|
208 |
void SiftMatchGL::SetDescriptors(int index, int num, const float* descriptors, int id) |
209 |
{ |
210 |
if(_initialized == 0) return; |
211 |
if (index > 1) index = 1; |
212 |
if (index < 0) index = 0; |
213 |
_have_loc[index] = 0;
|
214 |
|
215 |
//the same feature is already set
|
216 |
if(id !=-1 && id == _id_sift[index]) return ; |
217 |
_id_sift[index] = id; |
218 |
|
219 |
if(num > _max_sift) num = _max_sift;
|
220 |
|
221 |
sift_buffer.resize(num * 128 /4); |
222 |
unsigned char * pub = (unsigned char*) &sift_buffer[0]; |
223 |
for(int i = 0; i < 128 * num; ++i) |
224 |
{ |
225 |
pub[i] = int(512 * descriptors[i] + 0.5); |
226 |
} |
227 |
_num_sift[index] = num; |
228 |
int w = _sift_per_row * _pixel_per_sift;
|
229 |
int h = (num + _sift_per_row - 1)/ _sift_per_row; |
230 |
sift_buffer.resize(w * h * 4, 0); |
231 |
_texDes[index].SetImageSize(w, h); |
232 |
_texDes[index].BindTex(); |
233 |
if(_sift_num_stripe == 1) |
234 |
{ |
235 |
glTexSubImage2D(GlobalUtil::_texTarget, 0, 0, 0, w, h, GL_RGBA, GL_UNSIGNED_BYTE, &sift_buffer[0]); |
236 |
}else
|
237 |
{ |
238 |
for(int i = 0; i < _sift_num_stripe; ++i) |
239 |
{ |
240 |
int ws = _sift_per_stripe * _pixel_per_sift;
|
241 |
int x = i * ws;
|
242 |
int pos = i * ws * h * 4; |
243 |
glTexSubImage2D(GlobalUtil::_texTarget, 0, x, 0, ws, h, GL_RGBA, GL_UNSIGNED_BYTE, &sift_buffer[pos]); |
244 |
} |
245 |
} |
246 |
_texDes[index].UnbindTex(); |
247 |
} |
248 |
|
249 |
|
250 |
void SiftMatchGL::LoadSiftMatchShadersGLSL()
|
251 |
{ |
252 |
ProgramGLSL * program; |
253 |
char buffer[10240]; |
254 |
ostrstream out(buffer, 10240);
|
255 |
if(GlobalUtil::_IsNvidia)
|
256 |
out << "#pragma optionNV(ifcvt none)\n"
|
257 |
"#pragma optionNV(unroll all)\n";
|
258 |
|
259 |
out << "#define SIFT_PER_STRIPE " << _sift_per_stripe << ".0\n" |
260 |
"#define PIXEL_PER_SIFT " << _pixel_per_sift << "\n" |
261 |
"uniform sampler2DRect tex1, tex2; uniform vec2 size;\n"
|
262 |
"void main() \n"
|
263 |
"{\n"
|
264 |
<< " vec4 val = vec4(0, 0, 0, 0), data1, buf;\n"
|
265 |
" vec2 index = gl_FragCoord.yx; \n"
|
266 |
" vec2 stripe_size = size.xy * SIFT_PER_STRIPE;\n"
|
267 |
" vec2 temp_div1 = index / stripe_size;\n"
|
268 |
" vec2 stripe_index = floor(temp_div1);\n"
|
269 |
" index = floor(stripe_size * (temp_div1 - stripe_index));\n"
|
270 |
" vec2 temp_div2 = index * vec2(1.0 / float(SIFT_PER_STRIPE));\n"
|
271 |
" vec2 temp_floor2 = floor(temp_div2);\n"
|
272 |
" vec2 index_v = temp_floor2 + vec2(0.5);\n "
|
273 |
" vec2 index_h = vec2(SIFT_PER_STRIPE)* (temp_div2 - temp_floor2);\n"
|
274 |
" vec2 tx = (index_h + stripe_index * vec2(SIFT_PER_STRIPE))* vec2(PIXEL_PER_SIFT) + 0.5;\n"
|
275 |
" vec2 tpos1, tpos2; \n"
|
276 |
" vec4 tpos = vec4(tx, index_v);\n"
|
277 |
//////////////////////////////////////////////////////
|
278 |
" for(int i = 0; i < PIXEL_PER_SIFT; ++i){\n"
|
279 |
" buf = texture2DRect(tex2, tpos.yw);\n"
|
280 |
" data1 = texture2DRect(tex1, tpos.xz);\n"
|
281 |
" val += (data1 * buf);\n"
|
282 |
" tpos.xy = tpos.xy + vec2(1.0, 1.0);\n"
|
283 |
" }\n"
|
284 |
" const float factor = 0.248050689697265625; \n"
|
285 |
" gl_FragColor =vec4(dot(val, vec4(factor)), index, 0);\n"
|
286 |
"}"
|
287 |
<< '\0';
|
288 |
|
289 |
s_multiply = program= new ProgramGLSL(buffer);
|
290 |
|
291 |
_param_multiply_tex1 = glGetUniformLocation(*program, "tex1");
|
292 |
_param_multiply_tex2 = glGetUniformLocation(*program, "tex2");
|
293 |
_param_multiply_size = glGetUniformLocation(*program, "size");
|
294 |
|
295 |
out.seekp(ios::beg); |
296 |
if(GlobalUtil::_IsNvidia)
|
297 |
out << "#pragma optionNV(ifcvt none)\n"
|
298 |
"#pragma optionNV(unroll all)\n";
|
299 |
|
300 |
out << "#define SIFT_PER_STRIPE " << _sift_per_stripe << ".0\n" |
301 |
"#define PIXEL_PER_SIFT " << _pixel_per_sift << "\n" |
302 |
"uniform sampler2DRect tex1, tex2;\n"
|
303 |
"uniform sampler2DRect texL1;\n"
|
304 |
"uniform sampler2DRect texL2; \n"
|
305 |
"uniform mat3 H; \n"
|
306 |
"uniform mat3 F; \n"
|
307 |
"uniform vec4 size; \n"
|
308 |
"void main() \n"
|
309 |
"{\n"
|
310 |
<< " vec4 val = vec4(0, 0, 0, 0), data1, buf;\n"
|
311 |
" vec2 index = gl_FragCoord.yx; \n"
|
312 |
" vec2 stripe_size = size.xy * SIFT_PER_STRIPE;\n"
|
313 |
" vec2 temp_div1 = index / stripe_size;\n"
|
314 |
" vec2 stripe_index = floor(temp_div1);\n"
|
315 |
" index = floor(stripe_size * (temp_div1 - stripe_index));\n"
|
316 |
" vec2 temp_div2 = index * vec2(1.0/ float(SIFT_PER_STRIPE));\n"
|
317 |
" vec2 temp_floor2 = floor(temp_div2);\n"
|
318 |
" vec2 index_v = temp_floor2 + vec2(0.5);\n "
|
319 |
" vec2 index_h = vec2(SIFT_PER_STRIPE)* (temp_div2 - temp_floor2);\n"
|
320 |
|
321 |
//read feature location data
|
322 |
" vec4 tlpos = vec4((index_h + stripe_index * vec2(SIFT_PER_STRIPE)) + 0.5, index_v);\n"
|
323 |
" vec3 loc1 = vec3(texture2DRect(texL1, tlpos.xz).xw, 1.0);\n"
|
324 |
" vec3 loc2 = vec3(texture2DRect(texL2, tlpos.yw).xw, 1.0);\n"
|
325 |
|
326 |
//check the guiding homography
|
327 |
" vec3 hxloc1 = H* loc1;\n"
|
328 |
" vec2 diff = abs(loc2.xy- (hxloc1.xy/hxloc1.z));\n"
|
329 |
" float disth = max(diff.x, diff.y);\n"
|
330 |
" if(disth > size.z ) {gl_FragColor = vec4(0, index, 0); return;}\n"
|
331 |
|
332 |
//check the guiding fundamental
|
333 |
" vec3 fx1 = (F * loc1), ftx2 = (loc2 * F);\n"
|
334 |
" float x2tfx1 = dot(loc2, fx1);\n"
|
335 |
" vec4 temp = vec4(fx1.xy, ftx2.xy); \n"
|
336 |
" float sampson_error = (x2tfx1 * x2tfx1) / dot(temp, temp);\n"
|
337 |
" if(sampson_error > size.w) {gl_FragColor = vec4(0, index, 0); return;}\n"
|
338 |
|
339 |
//compare feature descriptor
|
340 |
" vec2 tx = (index_h + stripe_index * SIFT_PER_STRIPE)* vec2(PIXEL_PER_SIFT) + 0.5;\n"
|
341 |
" vec2 tpos1, tpos2; \n"
|
342 |
" vec4 tpos = vec4(tx, index_v);\n"
|
343 |
" for(int i = 0; i < PIXEL_PER_SIFT; ++i){\n"
|
344 |
" buf = texture2DRect(tex2, tpos.yw);\n"
|
345 |
" data1 = texture2DRect(tex1, tpos.xz);\n"
|
346 |
" val += data1 * buf;\n"
|
347 |
" tpos.xy = tpos.xy + vec2(1.0, 1.0);\n"
|
348 |
" }\n"
|
349 |
" const float factor = 0.248050689697265625; \n"
|
350 |
" gl_FragColor =vec4(dot(val, vec4(factor)), index, 0);\n"
|
351 |
"}"
|
352 |
<< '\0';
|
353 |
|
354 |
s_guided_mult = program= new ProgramGLSL(buffer);
|
355 |
|
356 |
_param_guided_mult_tex1 = glGetUniformLocation(*program, "tex1");
|
357 |
_param_guided_mult_tex2= glGetUniformLocation(*program, "tex2");
|
358 |
_param_guided_mult_texl1 = glGetUniformLocation(*program, "texL1");
|
359 |
_param_guided_mult_texl2 = glGetUniformLocation(*program, "texL2");
|
360 |
_param_guided_mult_h = glGetUniformLocation(*program, "H");
|
361 |
_param_guided_mult_f = glGetUniformLocation(*program, "F");
|
362 |
_param_guided_mult_param = glGetUniformLocation(*program, "size");
|
363 |
|
364 |
//row max
|
365 |
out.seekp(ios::beg); |
366 |
out << "#define BLOCK_WIDTH 16.0\n"
|
367 |
"uniform sampler2DRect tex; uniform vec3 param;\n"
|
368 |
"void main ()\n"
|
369 |
"{\n"
|
370 |
" float index = gl_FragCoord.x + floor(gl_FragCoord.y) * BLOCK_WIDTH; \n"
|
371 |
" vec2 bestv = vec2(-1.0); float imax = -1.0;\n"
|
372 |
" for(float i = 0.0; i < param.x; i ++){\n "
|
373 |
" float v = texture2DRect(tex, vec2(i + 0.5, index)).r; \n"
|
374 |
" imax = v > bestv.r ? i : imax; \n "
|
375 |
" bestv = v > bestv.r? vec2(v, bestv.r) : max(bestv, vec2(v));\n "
|
376 |
" }\n"
|
377 |
" bestv = acos(min(bestv, 1.0));\n"
|
378 |
" if(bestv.x >= param.y || bestv.x >= param.z * bestv.y) imax = -1.0;\n"
|
379 |
" gl_FragColor = vec4(imax, bestv, index);\n"
|
380 |
"}"
|
381 |
<< '\0';
|
382 |
s_row_max = program= new ProgramGLSL(buffer);
|
383 |
_param_rowmax_param = glGetUniformLocation(*program, "param");
|
384 |
|
385 |
out.seekp(ios::beg); |
386 |
out << "#define BLOCK_WIDTH 16.0\n"
|
387 |
"uniform sampler2DRect tex; uniform vec3 param;\n"
|
388 |
"void main ()\n"
|
389 |
"{\n"
|
390 |
" float index = gl_FragCoord.x + floor(gl_FragCoord.y) * BLOCK_WIDTH; \n"
|
391 |
" vec2 bestv = vec2(-1.0); float imax = -1.0;\n"
|
392 |
" for(float i = 0.0; i < param.x; i ++){\n "
|
393 |
" float v = texture2DRect(tex, vec2(index, i + 0.5)).r; \n"
|
394 |
" imax = (v > bestv.r)? i : imax; \n "
|
395 |
" bestv = v > bestv.r? vec2(v, bestv.r) : max(bestv, vec2(v));\n "
|
396 |
" }\n"
|
397 |
" bestv = acos(min(bestv, 1.0));\n"
|
398 |
" if(bestv.x >= param.y || bestv.x >= param.z * bestv.y) imax = -1.0;\n"
|
399 |
" gl_FragColor = vec4(imax, bestv, index);\n"
|
400 |
"}"
|
401 |
<< '\0';
|
402 |
s_col_max = program =new ProgramGLSL(buffer);
|
403 |
_param_colmax_param = glGetUniformLocation(*program, "param");
|
404 |
|
405 |
|
406 |
} |
407 |
|
408 |
int SiftMatchGL::GetGuidedSiftMatch(int max_match, int match_buffer[][2], float H[3][3], float F[3][3], |
409 |
float distmax, float ratiomax, float hdistmax, float fdistmax, int mbm) |
410 |
{ |
411 |
|
412 |
int dw = _num_sift[1]; |
413 |
int dh = _num_sift[0]; |
414 |
if(_initialized ==0) return 0; |
415 |
if(dw <= 0 || dh <=0) return 0; |
416 |
if(_have_loc[0] == 0 || _have_loc[1] == 0) return 0; |
417 |
|
418 |
FrameBufferObject fbo; |
419 |
glDrawBuffer(GL_COLOR_ATTACHMENT0_EXT); |
420 |
_texDot.SetImageSize(dw, dh); |
421 |
|
422 |
|
423 |
//data
|
424 |
_texDot.AttachToFBO(0);
|
425 |
_texDot.FitTexViewPort(); |
426 |
glActiveTexture(GL_TEXTURE0); |
427 |
_texDes[0].BindTex();
|
428 |
glActiveTexture(GL_TEXTURE1); |
429 |
_texDes[1].BindTex();
|
430 |
glActiveTexture(GL_TEXTURE2); |
431 |
_texLoc[0].BindTex();
|
432 |
glActiveTexture(GL_TEXTURE3); |
433 |
_texLoc[1].BindTex();
|
434 |
|
435 |
//multiply the descriptor matrices
|
436 |
s_guided_mult->UseProgram(); |
437 |
|
438 |
|
439 |
//set parameters glsl
|
440 |
float dot_param[4] = {(float)_texDes[0].GetDrawHeight(), (float) _texDes[1].GetDrawHeight(), hdistmax, fdistmax}; |
441 |
glUniform1i(_param_guided_mult_tex1, 0);
|
442 |
glUniform1i(_param_guided_mult_tex2, 1);
|
443 |
glUniform1i(_param_guided_mult_texl1, 2);
|
444 |
glUniform1i(_param_guided_mult_texl2, 3);
|
445 |
glUniformMatrix3fv(_param_guided_mult_h, 1, GL_TRUE, H[0]); |
446 |
glUniformMatrix3fv(_param_guided_mult_f, 1, GL_TRUE, F[0]); |
447 |
glUniform4fv(_param_guided_mult_param, 1, dot_param);
|
448 |
|
449 |
_texDot.DrawQuad(); |
450 |
|
451 |
GLTexImage::UnbindMultiTex(4);
|
452 |
|
453 |
return GetBestMatch(max_match, match_buffer, distmax, ratiomax, mbm);
|
454 |
} |
455 |
|
456 |
int SiftMatchGL::GetBestMatch(int max_match, int match_buffer[][2], float distmax, float ratiomax, int mbm) |
457 |
{ |
458 |
|
459 |
glActiveTexture(GL_TEXTURE0); |
460 |
_texDot.BindTex(); |
461 |
|
462 |
//readback buffer
|
463 |
sift_buffer.resize(_num_sift[0] + _num_sift[1] + 16); |
464 |
float * buffer1 = &sift_buffer[0], * buffer2 = &sift_buffer[_num_sift[0]]; |
465 |
|
466 |
//row max
|
467 |
_texMatch[0].AttachToFBO(0); |
468 |
_texMatch[0].SetImageSize(16, ( _num_sift[0] + 15) / 16); |
469 |
_texMatch[0].FitTexViewPort();
|
470 |
|
471 |
///set parameter glsl
|
472 |
s_row_max->UseProgram(); |
473 |
glUniform3f(_param_rowmax_param, (float)_num_sift[1], distmax, ratiomax); |
474 |
|
475 |
_texMatch[0].DrawQuad();
|
476 |
glReadPixels(0, 0, 16, (_num_sift[0] + 15)/16, GL_RED, GL_FLOAT, buffer1); |
477 |
|
478 |
//col max
|
479 |
if(mbm)
|
480 |
{ |
481 |
_texMatch[1].AttachToFBO(0); |
482 |
_texMatch[1].SetImageSize(16, (_num_sift[1] + 15) / 16); |
483 |
_texMatch[1].FitTexViewPort();
|
484 |
//set parameter glsl
|
485 |
s_col_max->UseProgram(); |
486 |
glUniform3f(_param_rowmax_param, (float)_num_sift[0], distmax, ratiomax); |
487 |
_texMatch[1].DrawQuad();
|
488 |
glReadPixels(0, 0, 16, (_num_sift[1] + 15) / 16, GL_RED, GL_FLOAT, buffer2); |
489 |
} |
490 |
|
491 |
|
492 |
//unload
|
493 |
glUseProgram(0);
|
494 |
|
495 |
GLTexImage::UnbindMultiTex(2);
|
496 |
GlobalUtil::CleanupOpenGL(); |
497 |
|
498 |
//write back the matches
|
499 |
int nmatch = 0, j ; |
500 |
for(int i = 0; i < _num_sift[0] && nmatch < max_match; ++i) |
501 |
{ |
502 |
j = int(buffer1[i]);
|
503 |
if( j>= 0 && (!mbm ||int(buffer2[j]) == i)) |
504 |
{ |
505 |
match_buffer[nmatch][0] = i;
|
506 |
match_buffer[nmatch][1] = j;
|
507 |
nmatch++; |
508 |
} |
509 |
} |
510 |
return nmatch;
|
511 |
} |
512 |
|
513 |
int SiftMatchGL::GetSiftMatch(int max_match, int match_buffer[][2], float distmax, float ratiomax, int mbm) |
514 |
{ |
515 |
int dw = _num_sift[1]; |
516 |
int dh = _num_sift[0]; |
517 |
if(_initialized ==0) return 0; |
518 |
if(dw <= 0 || dh <=0) return 0; |
519 |
|
520 |
FrameBufferObject fbo; |
521 |
glDrawBuffer(GL_COLOR_ATTACHMENT0_EXT); |
522 |
_texDot.SetImageSize(dw, dh); |
523 |
|
524 |
//data
|
525 |
_texDot.AttachToFBO(0);
|
526 |
_texDot.FitTexViewPort(); |
527 |
glActiveTexture(GL_TEXTURE0); |
528 |
_texDes[0].BindTex();
|
529 |
glActiveTexture(GL_TEXTURE1); |
530 |
_texDes[1].BindTex();
|
531 |
|
532 |
//////////////////
|
533 |
//multiply the descriptor matrices
|
534 |
s_multiply->UseProgram(); |
535 |
//set parameters
|
536 |
float heights[2] = {(float)_texDes[0].GetDrawHeight(), (float)_texDes[1].GetDrawHeight()}; |
537 |
|
538 |
glUniform1i(_param_multiply_tex1, 0);
|
539 |
glUniform1i(_param_multiply_tex2 , 1);
|
540 |
glUniform2fv(_param_multiply_size, 1, heights);
|
541 |
|
542 |
_texDot.DrawQuad(); |
543 |
|
544 |
glActiveTexture(GL_TEXTURE1); |
545 |
glBindTexture(GlobalUtil::_texTarget, 0);
|
546 |
|
547 |
return GetBestMatch(max_match, match_buffer, distmax, ratiomax, mbm);
|
548 |
} |
549 |
|
550 |
|
551 |
int SiftMatchGPU::_CreateContextGL()
|
552 |
{ |
553 |
//Create an OpenGL Context?
|
554 |
if (__language >= SIFTMATCH_CUDA) {}
|
555 |
else if(!GlobalUtil::CreateWindowEZ()) |
556 |
{ |
557 |
#if CUDA_SIFTGPU_ENABLED
|
558 |
__language = SIFTMATCH_CUDA; |
559 |
#else
|
560 |
return 0; |
561 |
#endif
|
562 |
} |
563 |
return VerifyContextGL();
|
564 |
} |
565 |
|
566 |
|
567 |
int SiftMatchGPU::_VerifyContextGL()
|
568 |
{ |
569 |
if(__matcher) return GlobalUtil::_GoodOpenGL; |
570 |
|
571 |
#ifdef CUDA_SIFTGPU_ENABLED
|
572 |
|
573 |
if(__language >= SIFTMATCH_CUDA) {}
|
574 |
else if(__language == SIFTMATCH_SAME_AS_SIFTGPU && GlobalUtil::_UseCUDA){} |
575 |
else GlobalUtil::InitGLParam(0); |
576 |
if(GlobalUtil::_GoodOpenGL == 0) __language = SIFTMATCH_CUDA; |
577 |
|
578 |
if(((__language == SIFTMATCH_SAME_AS_SIFTGPU && GlobalUtil::_UseCUDA) || __language >= SIFTMATCH_CUDA)
|
579 |
&& SiftMatchCU::CheckCudaDevice (GlobalUtil::_DeviceIndex)) |
580 |
{ |
581 |
__language = SIFTMATCH_CUDA; |
582 |
__matcher = new SiftMatchCU(__max_sift);
|
583 |
}else
|
584 |
#else
|
585 |
if((__language == SIFTMATCH_SAME_AS_SIFTGPU && GlobalUtil::_UseCUDA) || __language >= SIFTMATCH_CUDA)
|
586 |
{ |
587 |
std::cerr << "---------------------------------------------------------------------------\n"
|
588 |
<< "CUDA not supported in this binary! To enable it, please use SiftGPU_CUDA_Enable\n"
|
589 |
<< "Project for VS2005+ or set siftgpu_enable_cuda to 1 in makefile\n"
|
590 |
<< "----------------------------------------------------------------------------\n";
|
591 |
} |
592 |
#endif
|
593 |
{ |
594 |
__language = SIFTMATCH_GLSL; |
595 |
__matcher = new SiftMatchGL(__max_sift, 1); |
596 |
} |
597 |
|
598 |
if(GlobalUtil::_verbose)
|
599 |
std::cout << "[SiftMatchGPU]: " << (__language == SIFTMATCH_CUDA? "CUDA" : "GLSL") <<"\n\n"; |
600 |
|
601 |
__matcher->InitSiftMatch(); |
602 |
return GlobalUtil::_GoodOpenGL;
|
603 |
} |
604 |
|
605 |
void* SiftMatchGPU::operator new (size_t size){ |
606 |
void * p = malloc(size);
|
607 |
if (p == 0) |
608 |
{ |
609 |
const std::bad_alloc ba;
|
610 |
throw ba;
|
611 |
} |
612 |
return p;
|
613 |
} |
614 |
|
615 |
|
616 |
SiftMatchGPU::SiftMatchGPU(int max_sift)
|
617 |
{ |
618 |
__max_sift = max(max_sift, 1024);
|
619 |
__language = 0;
|
620 |
__matcher = NULL;
|
621 |
} |
622 |
|
623 |
void SiftMatchGPU::SetLanguage(int language) |
624 |
{ |
625 |
if(__matcher) return; |
626 |
////////////////////////
|
627 |
#ifdef CUDA_SIFTGPU_ENABLED
|
628 |
if(language >= SIFTMATCH_CUDA) GlobalUtil::_DeviceIndex = language - SIFTMATCH_CUDA;
|
629 |
#endif
|
630 |
__language = language > SIFTMATCH_CUDA ? SIFTMATCH_CUDA : language; |
631 |
} |
632 |
|
633 |
void SiftMatchGPU::SetDeviceParam(int argc, char**argv) |
634 |
{ |
635 |
if(__matcher) return; |
636 |
GlobalUtil::SetDeviceParam(argc, argv); |
637 |
} |
638 |
|
639 |
void SiftMatchGPU::SetMaxSift(int max_sift) |
640 |
{ |
641 |
if(__matcher) __matcher->SetMaxSift(max(128, max_sift)); |
642 |
else __max_sift = max(128, max_sift); |
643 |
} |
644 |
|
645 |
SiftMatchGPU::~SiftMatchGPU() |
646 |
{ |
647 |
if(__matcher) delete __matcher; |
648 |
} |
649 |
|
650 |
void SiftMatchGPU::SetDescriptors(int index, int num, const unsigned char* descriptors, int id) |
651 |
{ |
652 |
__matcher->SetDescriptors(index, num, descriptors, id); |
653 |
} |
654 |
|
655 |
void SiftMatchGPU::SetDescriptors(int index, int num, const float* descriptors, int id) |
656 |
{ |
657 |
__matcher->SetDescriptors(index, num, descriptors, id); |
658 |
} |
659 |
|
660 |
void SiftMatchGPU::SetFeautreLocation(int index, const float* locations, int gap) |
661 |
{ |
662 |
__matcher->SetFeautreLocation(index, locations, gap); |
663 |
|
664 |
} |
665 |
int SiftMatchGPU::GetGuidedSiftMatch(int max_match, int match_buffer[][2], float H[3][3], float F[3][3], |
666 |
float distmax, float ratiomax, float hdistmax, float fdistmax, int mutual_best_match) |
667 |
{ |
668 |
if(H == NULL && F == NULL) |
669 |
{ |
670 |
return __matcher->GetSiftMatch(max_match, match_buffer, distmax, ratiomax, mutual_best_match);
|
671 |
}else
|
672 |
{ |
673 |
float Z[3][3] = {{1, 0, 0}, {0, 1, 0}, {0, 0, 1}}, ti = (1.0e+20F); |
674 |
|
675 |
return __matcher->GetGuidedSiftMatch(max_match, match_buffer, H? H : Z, F? F : Z,
|
676 |
distmax, ratiomax, H? hdistmax: ti, F? fdistmax: ti, mutual_best_match); |
677 |
} |
678 |
} |
679 |
|
680 |
int SiftMatchGPU::GetSiftMatch(int max_match, int match_buffer[][2], float distmax, float ratiomax, int mutual_best_match) |
681 |
{ |
682 |
return __matcher->GetSiftMatch(max_match, match_buffer, distmax, ratiomax, mutual_best_match);
|
683 |
} |
684 |
|
685 |
SiftMatchGPU* CreateNewSiftMatchGPU(int max_sift)
|
686 |
{ |
687 |
return new SiftMatchGPU(max_sift); |
688 |
} |
689 |
|