2 #include "_latentsvm.h"
3 #include "_lsvm_resizeimg.h"
6 #define max(a,b) (((a) > (b)) ? (a) : (b))
10 #define min(a,b) (((a) < (b)) ? (a) : (b))
14 // Getting feature map for the selected subimage
17 // int getFeatureMaps(const IplImage * image, const int k, featureMap **map);
19 // image - selected subimage
26 int getFeatureMaps(const IplImage* image, const int k, CvLSVMFeatureMap **map)
29 int p, px, stringSize;
30 int height, width, numChannels;
31 int i, j, kk, c, ii, jj, d;
32 float * datadx, * datady;
35 float magnitude, x, y, tx, ty;
41 float kernel[3] = {-1.f, 0.f, 1.f};
42 CvMat kernel_dx = cvMat(1, 3, CV_32F, kernel);
43 CvMat kernel_dy = cvMat(3, 1, CV_32F, kernel);
48 float boundary_x[NUM_SECTOR + 1];
49 float boundary_y[NUM_SECTOR + 1];
53 height = image->height;
54 width = image->width ;
56 numChannels = image->nChannels;
58 dx = cvCreateImage(cvSize(image->width, image->height),
60 dy = cvCreateImage(cvSize(image->width, image->height),
67 stringSize = sizeX * p;
68 allocFeatureMapObject(map, sizeX, sizeY, p);
70 cvFilter2D(image, dx, &kernel_dx, cvPoint(-1, 0));
71 cvFilter2D(image, dy, &kernel_dy, cvPoint(0, -1));
74 for(i = 0; i <= NUM_SECTOR; i++)
76 arg_vector = ( (float) i ) * ( (float)(PI) / (float)(NUM_SECTOR) );
77 boundary_x[i] = cosf(arg_vector);
78 boundary_y[i] = sinf(arg_vector);
79 }/*for(i = 0; i <= NUM_SECTOR; i++) */
81 r = (float *)malloc( sizeof(float) * (width * height));
82 alfa = (int *)malloc( sizeof(int ) * (width * height * 2));
84 for(j = 1; j < height - 1; j++)
86 datadx = (float*)(dx->imageData + dx->widthStep * j);
87 datady = (float*)(dy->imageData + dy->widthStep * j);
88 for(i = 1; i < width - 1; i++)
91 x = (datadx[i * numChannels + c]);
92 y = (datady[i * numChannels + c]);
94 r[j * width + i] =sqrtf(x * x + y * y);
95 for(ch = 1; ch < numChannels; ch++)
97 tx = (datadx[i * numChannels + ch]);
98 ty = (datady[i * numChannels + ch]);
99 magnitude = sqrtf(tx * tx + ty * ty);
100 if(magnitude > r[j * width + i])
102 r[j * width + i] = magnitude;
107 }/*for(ch = 1; ch < numChannels; ch++)*/
109 max = boundary_x[0] * x + boundary_y[0] * y;
111 for (kk = 0; kk < NUM_SECTOR; kk++)
113 dotProd = boundary_x[kk] * x + boundary_y[kk] * y;
124 maxi = kk + NUM_SECTOR;
128 alfa[j * width * 2 + i * 2 ] = maxi % NUM_SECTOR;
129 alfa[j * width * 2 + i * 2 + 1] = maxi;
130 }/*for(i = 0; i < width; i++)*/
131 }/*for(j = 0; j < height; j++)*/
133 nearest = (int *)malloc(sizeof(int ) * k);
134 w = (float*)malloc(sizeof(float) * (k * 2));
136 for(i = 0; i < k / 2; i++)
139 }/*for(i = 0; i < k / 2; i++)*/
140 for(i = k / 2; i < k; i++)
143 }/*for(i = k / 2; i < k; i++)*/
145 for(j = 0; j < k / 2; j++)
147 b_x = k / 2 + j + 0.5f;
148 a_x = k / 2 - j - 0.5f;
149 w[j * 2 ] = 1.0f/a_x * ((a_x * b_x) / ( a_x + b_x));
150 w[j * 2 + 1] = 1.0f/b_x * ((a_x * b_x) / ( a_x + b_x));
151 }/*for(j = 0; j < k / 2; j++)*/
152 for(j = k / 2; j < k; j++)
154 a_x = j - k / 2 + 0.5f;
155 b_x =-j + k / 2 - 0.5f + k;
156 w[j * 2 ] = 1.0f/a_x * ((a_x * b_x) / ( a_x + b_x));
157 w[j * 2 + 1] = 1.0f/b_x * ((a_x * b_x) / ( a_x + b_x));
158 }/*for(j = k / 2; j < k; j++)*/
161 for(i = 0; i < sizeY; i++)
163 for(j = 0; j < sizeX; j++)
165 for(ii = 0; ii < k; ii++)
167 for(jj = 0; jj < k; jj++)
169 if ((i * k + ii > 0) &&
170 (i * k + ii < height - 1) &&
172 (j * k + jj < width - 1))
174 d = (k * i + ii) * width + (j * k + jj);
175 (*map)->map[ i * stringSize + j * (*map)->numFeatures + alfa[d * 2 ]] +=
176 r[d] * w[ii * 2] * w[jj * 2];
177 (*map)->map[ i * stringSize + j * (*map)->numFeatures + alfa[d * 2 + 1] + NUM_SECTOR] +=
178 r[d] * w[ii * 2] * w[jj * 2];
179 if ((i + nearest[ii] >= 0) &&
180 (i + nearest[ii] <= sizeY - 1))
182 (*map)->map[(i + nearest[ii]) * stringSize + j * (*map)->numFeatures + alfa[d * 2 ] ] +=
183 r[d] * w[ii * 2 + 1] * w[jj * 2 ];
184 (*map)->map[(i + nearest[ii]) * stringSize + j * (*map)->numFeatures + alfa[d * 2 + 1] + NUM_SECTOR] +=
185 r[d] * w[ii * 2 + 1] * w[jj * 2 ];
187 if ((j + nearest[jj] >= 0) &&
188 (j + nearest[jj] <= sizeX - 1))
190 (*map)->map[i * stringSize + (j + nearest[jj]) * (*map)->numFeatures + alfa[d * 2 ] ] +=
191 r[d] * w[ii * 2] * w[jj * 2 + 1];
192 (*map)->map[i * stringSize + (j + nearest[jj]) * (*map)->numFeatures + alfa[d * 2 + 1] + NUM_SECTOR] +=
193 r[d] * w[ii * 2] * w[jj * 2 + 1];
195 if ((i + nearest[ii] >= 0) &&
196 (i + nearest[ii] <= sizeY - 1) &&
197 (j + nearest[jj] >= 0) &&
198 (j + nearest[jj] <= sizeX - 1))
200 (*map)->map[(i + nearest[ii]) * stringSize + (j + nearest[jj]) * (*map)->numFeatures + alfa[d * 2 ] ] +=
201 r[d] * w[ii * 2 + 1] * w[jj * 2 + 1];
202 (*map)->map[(i + nearest[ii]) * stringSize + (j + nearest[jj]) * (*map)->numFeatures + alfa[d * 2 + 1] + NUM_SECTOR] +=
203 r[d] * w[ii * 2 + 1] * w[jj * 2 + 1];
206 }/*for(jj = 0; jj < k; jj++)*/
207 }/*for(ii = 0; ii < k; ii++)*/
208 }/*for(j = 1; j < sizeX - 1; j++)*/
209 }/*for(i = 1; i < sizeY - 1; i++)*/
221 return LATENT_SVM_OK;
225 // Feature map Normalization and Truncation
228 // int normalizeAndTruncate(featureMap *map, const float alfa);
231 // alfa - truncation threshold
233 // map - truncated and normalized feature map
237 int normalizeAndTruncate(CvLSVMFeatureMap *map, const float alfa)
240 int sizeX, sizeY, p, pos, pp, xp, pos1, pos2;
241 float * partOfNorm; // norm of C(i, j)
247 partOfNorm = (float *)malloc (sizeof(float) * (sizeX * sizeY));
251 pp = NUM_SECTOR * 12;
253 for(i = 0; i < sizeX * sizeY; i++)
256 pos = i * map->numFeatures;
257 for(j = 0; j < p; j++)
259 valOfNorm += map->map[pos + j] * map->map[pos + j];
260 }/*for(j = 0; j < p; j++)*/
261 partOfNorm[i] = valOfNorm;
262 }/*for(i = 0; i < sizeX * sizeY; i++)*/
267 newData = (float *)malloc (sizeof(float) * (sizeX * sizeY * pp));
269 for(i = 1; i <= sizeY; i++)
271 for(j = 1; j <= sizeX; j++)
274 partOfNorm[(i )*(sizeX + 2) + (j )] +
275 partOfNorm[(i )*(sizeX + 2) + (j + 1)] +
276 partOfNorm[(i + 1)*(sizeX + 2) + (j )] +
277 partOfNorm[(i + 1)*(sizeX + 2) + (j + 1)]) + FLT_EPSILON;
278 pos1 = (i ) * (sizeX + 2) * xp + (j ) * xp;
279 pos2 = (i-1) * (sizeX ) * pp + (j-1) * pp;
280 for(ii = 0; ii < p; ii++)
282 newData[pos2 + ii ] = map->map[pos1 + ii ] / valOfNorm;
283 }/*for(ii = 0; ii < p; ii++)*/
284 for(ii = 0; ii < 2 * p; ii++)
286 newData[pos2 + ii + p * 4] = map->map[pos1 + ii + p] / valOfNorm;
287 }/*for(ii = 0; ii < 2 * p; ii++)*/
289 partOfNorm[(i )*(sizeX + 2) + (j )] +
290 partOfNorm[(i )*(sizeX + 2) + (j + 1)] +
291 partOfNorm[(i - 1)*(sizeX + 2) + (j )] +
292 partOfNorm[(i - 1)*(sizeX + 2) + (j + 1)]) + FLT_EPSILON;
293 for(ii = 0; ii < p; ii++)
295 newData[pos2 + ii + p ] = map->map[pos1 + ii ] / valOfNorm;
296 }/*for(ii = 0; ii < p; ii++)*/
297 for(ii = 0; ii < 2 * p; ii++)
299 newData[pos2 + ii + p * 6] = map->map[pos1 + ii + p] / valOfNorm;
300 }/*for(ii = 0; ii < 2 * p; ii++)*/
302 partOfNorm[(i )*(sizeX + 2) + (j )] +
303 partOfNorm[(i )*(sizeX + 2) + (j - 1)] +
304 partOfNorm[(i + 1)*(sizeX + 2) + (j )] +
305 partOfNorm[(i + 1)*(sizeX + 2) + (j - 1)]) + FLT_EPSILON;
306 for(ii = 0; ii < p; ii++)
308 newData[pos2 + ii + p * 2] = map->map[pos1 + ii ] / valOfNorm;
309 }/*for(ii = 0; ii < p; ii++)*/
310 for(ii = 0; ii < 2 * p; ii++)
312 newData[pos2 + ii + p * 8] = map->map[pos1 + ii + p] / valOfNorm;
313 }/*for(ii = 0; ii < 2 * p; ii++)*/
315 partOfNorm[(i )*(sizeX + 2) + (j )] +
316 partOfNorm[(i )*(sizeX + 2) + (j - 1)] +
317 partOfNorm[(i - 1)*(sizeX + 2) + (j )] +
318 partOfNorm[(i - 1)*(sizeX + 2) + (j - 1)]) + FLT_EPSILON;
319 for(ii = 0; ii < p; ii++)
321 newData[pos2 + ii + p * 3 ] = map->map[pos1 + ii ] / valOfNorm;
322 }/*for(ii = 0; ii < p; ii++)*/
323 for(ii = 0; ii < 2 * p; ii++)
325 newData[pos2 + ii + p * 10] = map->map[pos1 + ii + p] / valOfNorm;
326 }/*for(ii = 0; ii < 2 * p; ii++)*/
327 }/*for(j = 1; j <= sizeX; j++)*/
328 }/*for(i = 1; i <= sizeY; i++)*/
330 for(i = 0; i < sizeX * sizeY * pp; i++)
332 if(newData [i] > alfa) newData [i] = alfa;
333 }/*for(i = 0; i < sizeX * sizeY * pp; i++)*/
336 map->numFeatures = pp;
345 return LATENT_SVM_OK;
349 // Feature map reduction
350 // In each cell we reduce dimension of the feature vector
351 // according to original paper special procedure
354 // int PCAFeatureMaps(featureMap *map)
362 int PCAFeatureMaps(CvLSVMFeatureMap *map)
365 int sizeX, sizeY, p, pp, xp, yp, pos1, pos2;
372 p = map->numFeatures;
373 pp = NUM_SECTOR * 3 + 4;
377 nx = 1.0f / sqrtf((float)(xp * 2));
378 ny = 1.0f / sqrtf((float)(yp ));
380 newData = (float *)malloc (sizeof(float) * (sizeX * sizeY * pp));
382 for(i = 0; i < sizeY; i++)
384 for(j = 0; j < sizeX; j++)
386 pos1 = ((i)*sizeX + j)*p;
387 pos2 = ((i)*sizeX + j)*pp;
389 for(jj = 0; jj < xp * 2; jj++)
392 for(ii = 0; ii < yp; ii++)
394 val += map->map[pos1 + yp * xp + ii * xp * 2 + jj];
395 }/*for(ii = 0; ii < yp; ii++)*/
396 newData[pos2 + k] = val * ny;
398 }/*for(jj = 0; jj < xp * 2; jj++)*/
399 for(jj = 0; jj < xp; jj++)
402 for(ii = 0; ii < yp; ii++)
404 val += map->map[pos1 + ii * xp + jj];
405 }/*for(ii = 0; ii < yp; ii++)*/
406 newData[pos2 + k] = val * ny;
408 }/*for(jj = 0; jj < xp; jj++)*/
409 for(ii = 0; ii < yp; ii++)
412 for(jj = 0; jj < 2 * xp; jj++)
414 val += map->map[pos1 + yp * xp + ii * xp * 2 + jj];
415 }/*for(jj = 0; jj < xp; jj++)*/
416 newData[pos2 + k] = val * nx;
418 } /*for(ii = 0; ii < yp; ii++)*/
419 }/*for(j = 0; j < sizeX; j++)*/
420 }/*for(i = 0; i < sizeY; i++)*/
423 map->numFeatures = pp;
429 return LATENT_SVM_OK;
433 static int getPathOfFeaturePyramid(IplImage * image,
434 float step, int numStep, int startIndex,
435 int sideLength, CvLSVMFeaturePyramid **maps)
437 CvLSVMFeatureMap *map;
442 for(i = 0; i < numStep; i++)
444 scale = 1.0f / powf(step, (float)i);
445 scaleTmp = resize_opencv (image, scale);
446 getFeatureMaps(scaleTmp, sideLength, &map);
447 normalizeAndTruncate(map, VAL_OF_TRUNCATE);
449 (*maps)->pyramid[startIndex + i] = map;
450 cvReleaseImage(&scaleTmp);
451 }/*for(i = 0; i < numStep; i++)*/
452 return LATENT_SVM_OK;
456 // Getting feature pyramid
459 // int getFeaturePyramid(IplImage * image, const filterObject **all_F,
461 const int lambda, const int k,
462 const int startX, const int startY,
463 const int W, const int H, featurePyramid **maps);
467 // maps - feature maps for all levels
471 int getFeaturePyramid(IplImage * image, CvLSVMFeaturePyramid **maps)
479 if(image->depth == IPL_DEPTH_32F)
485 imgResize = cvCreateImage(cvSize(image->width , image->height) ,
487 cvConvert(image, imgResize);
490 W = imgResize->width;
491 H = imgResize->height;
493 step = powf(2.0f, 1.0f / ((float)LAMBDA));
494 maxNumCells = W / SIDE_LENGTH;
495 if( maxNumCells > H / SIDE_LENGTH )
497 maxNumCells = H / SIDE_LENGTH;
499 numStep = (int)(logf((float) maxNumCells / (5.0f)) / logf( step )) + 1;
501 allocFeaturePyramidObject(maps, numStep + LAMBDA);
503 getPathOfFeaturePyramid(imgResize, step , LAMBDA, 0,
504 SIDE_LENGTH / 2, maps);
505 getPathOfFeaturePyramid(imgResize, step, numStep, LAMBDA,
508 if(image->depth != IPL_DEPTH_32F)
510 cvReleaseImage(&imgResize);
513 return LATENT_SVM_OK;