Commit | Line | Data |
---|---|---|
80f575fc DM |
1 | /* |
2 | * localmotion2transform.c | |
3 | * | |
4 | * Copyright (C) Georg Martius - January 2013 | |
5 | * georg dot martius at web dot de | |
6 | * | |
7 | * This file is part of vid.stab video stabilization library | |
8 | * | |
9 | * vid.stab is free software; you can redistribute it and/or modify | |
10 | * it under the terms of the GNU General Public License, | |
11 | * as published by the Free Software Foundation; either version 2, or | |
12 | * (at your option) any later version. | |
13 | * | |
14 | * vid.stab is distributed in the hope that it will be useful, | |
15 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
16 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
17 | * GNU General Public License for more details. | |
18 | * | |
19 | * You should have received a copy of the GNU General Public License | |
20 | * along with GNU Make; see the file COPYING. If not, write to | |
21 | * the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA. | |
22 | * | |
23 | */ | |
24 | ||
25 | #include "localmotion2transform.h" | |
26 | #include "transformtype_operations.h" | |
27 | #include <assert.h> | |
28 | #include <string.h> | |
29 | ||
30 | /* #include <sys/time.h> */ | |
31 | /* long timeOfDayinMS() { */ | |
32 | /* struct timeval t; */ | |
33 | /* gettimeofday(&t, 0); */ | |
34 | /* return t.tv_sec*1000 + t.tv_usec/1000; */ | |
35 | /* } */ | |
36 | ||
37 | int vsLocalmotions2Transforms(VSTransformData* td, | |
38 | const VSManyLocalMotions* motions, | |
39 | VSTransformations* trans ){ | |
40 | int len = vs_vector_size(motions); | |
41 | assert(trans->len==0 && trans->ts == 0); | |
42 | trans->ts = vs_malloc(sizeof(VSTransform)*len ); | |
43 | /* long start= timeOfDayinMS(); */ | |
44 | FILE *f=0; | |
45 | if(td->conf.storeTransforms){ | |
46 | f = fopen("global_motions.trf","w"); | |
47 | } | |
48 | ||
49 | if(td->conf.simpleMotionCalculation==0){ | |
50 | for(int i=0; i< vs_vector_size(motions); i++) { | |
51 | trans->ts[i]=vsMotionsToTransform(td,VSMLMGet(motions,i), f); | |
52 | } | |
53 | }else{ | |
54 | for(int i=0; i< vs_vector_size(motions); i++) { | |
55 | trans->ts[i]=vsSimpleMotionsToTransform(td->fiSrc, td->conf.modName,VSMLMGet(motions,i)); | |
56 | } | |
57 | } | |
58 | trans->len=len; | |
59 | ||
60 | /* long end = timeOfDayinMS(); */ | |
61 | /* vs_log_info(td->conf.modName, "Localmotions2Transform (%i) with %i frames took %i ms\n", */ | |
62 | /* td->conf.simpleMotionCalculation, len, end-start); */ | |
63 | if(f) fclose(f); | |
64 | return VS_OK; | |
65 | } | |
66 | ||
67 | VSArray vsTransformToArray(const VSTransform* t){ | |
68 | VSArray a = vs_array_new(4); | |
69 | a.dat[0] = t->x; | |
70 | a.dat[1] = t->y; | |
71 | a.dat[2] = t->alpha; | |
72 | a.dat[3] = t->zoom; | |
73 | return a; | |
74 | } | |
75 | ||
76 | VSTransform vsArrayToTransform(VSArray a){ | |
77 | return new_transform(a.dat[0],a.dat[1],a.dat[2],a.dat[3],0,0,0); | |
78 | } | |
79 | ||
80 | struct VSGradientDat { | |
81 | VSTransformData* td; | |
82 | const LocalMotions* motions; | |
83 | VSArray missmatches; // if negative then local motion is ignored | |
84 | }; | |
85 | ||
86 | double calcTransformQuality(VSArray params, void* dat){ | |
87 | struct VSGradientDat* gd= (struct VSGradientDat*) dat; | |
88 | const LocalMotions* motions = gd->motions; | |
89 | int num_motions=vs_vector_size(motions); | |
90 | VSTransform t = vsArrayToTransform(params); | |
91 | double error=0; | |
92 | ||
93 | PreparedTransform pt= prepare_transform(&t, &gd->td->fiSrc); | |
94 | int num = 1; // we start with 1 to avoid div by zero | |
95 | for (int i = 0; i < num_motions; i++) { | |
96 | if(gd->missmatches.dat[i]>=0){ | |
97 | LocalMotion* m = LMGet(motions,i); | |
98 | double vx,vy; | |
99 | transform_vec_double(&vx, &vy, &pt, (Vec*)&m->f); | |
100 | vx -= m->f.x; vy -= m->f.y; | |
101 | double e = sqr(vx - m->v.x) + sqr(vy - m->v.y); | |
102 | gd->missmatches.dat[i]=e; | |
103 | error += e; | |
104 | num++; | |
105 | } | |
106 | } | |
107 | // 1 pixel translation missmatch is roughly (with size 500): | |
108 | // alpha=0.11 (degree), zoom=0.2; The zoom is however often much larger, so less penalty. | |
109 | return error/num + fabs(t.alpha)/5.0 + fabs(t.zoom)/500.0; | |
110 | } | |
111 | ||
112 | double intMean(const int* ds, int len) { | |
113 | double sum=0; | |
114 | for (int i = 0; i < len; i++) sum += ds[i]; | |
115 | return sum / len; | |
116 | } | |
117 | ||
118 | // only calcates means transform to initialise gradient descent | |
119 | VSTransform meanMotions(VSTransformData* td, const LocalMotions* motions){ | |
120 | int len = vs_vector_size(motions); | |
121 | int* xs = localmotions_getx(motions); | |
122 | int* ys = localmotions_gety(motions); | |
123 | VSTransform t = null_transform(); | |
124 | if(motions==0 || len==0) { | |
125 | t.extra = 1; // prob. blank frame or too low contrast, ignore later | |
126 | return t; | |
127 | } | |
128 | t.x = intMean(xs,len); | |
129 | t.y = intMean(ys,len); | |
130 | vs_free(xs); | |
131 | vs_free(ys); | |
132 | return t; | |
133 | } | |
134 | ||
135 | /* Disables those fields (mask = -1) whose (miss)quality is high. | |
136 | @param mask: fields masks (<0 means disabled) | |
137 | @param missqualities: measure for each field (larger is worse) | |
138 | @param stddevs: x standard deviations to exclude | |
139 | Both array have to be of the same length. | |
140 | @return number of disabled fields | |
141 | */ | |
142 | int disableFields(VSArray mask, VSArray missqualities, double stddevs){ | |
143 | assert(mask.len == missqualities.len); | |
144 | // first we throw away those fields that match badely (during motion detection) | |
145 | double mu = mean(missqualities.dat, missqualities.len); | |
146 | double sigma = stddev(missqualities.dat, missqualities.len, mu); | |
147 | double thresh = mu + stddevs * sigma; | |
148 | int cnt=0; | |
149 | for(int i=0; i< mask.len; i++){ | |
150 | if(missqualities.dat[i]>thresh){ | |
151 | mask.dat[i]=-1.0; // disable field | |
152 | cnt++; | |
153 | } | |
154 | } | |
155 | return cnt; | |
156 | } | |
157 | ||
158 | VSTransform vsMotionsToTransform(VSTransformData* td, | |
159 | const LocalMotions* motions, | |
160 | FILE* f){ | |
161 | VSTransform t = meanMotions(td, motions); | |
162 | if(motions==0 || vs_vector_size(motions)==0){ | |
163 | if (f) fprintf(f,"0 0 0 0 0 %i\n# no fields\n", t.extra); | |
164 | return t; | |
165 | } | |
166 | VSArray missmatches = vs_array_new(vs_vector_size(motions)); | |
167 | VSArray params = vsTransformToArray(&t); | |
168 | double residual; | |
169 | struct VSGradientDat dat; | |
170 | dat.motions = motions; | |
171 | dat.td = td; | |
172 | dat.missmatches = missmatches; | |
173 | ||
174 | // first we throw away those fields that match badely (during motion detection) | |
175 | VSArray matchQualities = localmotionsGetMatch(motions); | |
176 | int dis1=disableFields(missmatches, matchQualities, 1.5); | |
177 | vs_array_free(matchQualities); | |
178 | ||
179 | VSArray result; | |
180 | double ss[] = {0.2, 0.2, 0.00005, 0.1}; | |
181 | int k; | |
182 | int dis2=0; | |
183 | for(k=0; k<3; k++){ | |
184 | // optimize params to minimize transform quality (12 steps per dimension) | |
185 | result = vsGradientDescent(calcTransformQuality, params, &dat, | |
186 | 16, vs_array(ss,4), 0.01, &residual); | |
187 | vs_array_free(params); | |
188 | // now we need to ignore the fields that don't fit well (e.g. moving objects) | |
189 | // cut off everthing above 1 std. dev. for skewed distributions | |
190 | // this will cut off the tail | |
191 | // do this only two times (3 gradient optimizations in total) | |
192 | if((k==0 && residual>0.1) || (k==1 && residual>20)){ | |
193 | dis2 += disableFields(missmatches, missmatches, 1.0); | |
194 | params = result; | |
195 | } else break; | |
196 | } | |
197 | ||
198 | if(td->conf.verbose & VS_DEBUG) | |
199 | vs_log_info(td->conf.modName, "disabled (%i+%i)/%i,\tresidual: %f (%i)\n", | |
200 | dis1, dis2, vs_vector_size(motions), residual, k+1); | |
201 | t = vsArrayToTransform(result); | |
202 | vs_array_free(result); | |
203 | vs_array_free(missmatches); | |
204 | // check if sufficiently good match was achieved: | |
205 | if(residual>100){ // test threshold. | |
206 | t.extra=1; | |
207 | } | |
208 | if(f){ | |
209 | fprintf(f,"0 %f %f %f %f %i\n#\t\t\t\t\t %f %i\n", t.x, t.y, t.alpha, t.zoom, t.extra, | |
210 | residual, k + 1); | |
211 | } | |
212 | if(!td->conf.smoothZoom) | |
213 | t.zoom=0; | |
214 | return t; | |
215 | } | |
216 | ||
217 | ||
218 | ||
219 | /* n-dimensional general purpose gradient descent algorithm */ | |
220 | VSArray vsGradientDescent(double (*eval)(VSArray, void*), | |
221 | VSArray params, void* dat, | |
222 | int N, VSArray stepsizes, double threshold, double* residual){ | |
223 | int dim=params.len; | |
224 | double v = eval(params, dat); | |
225 | VSArray x = vs_array_copy(params); | |
226 | VSArray grad = vs_array_new(dim); | |
227 | assert(stepsizes.len == params.len); | |
228 | for(int i=0; i< N*dim && v > threshold; i++){ | |
229 | int k=i%dim; | |
230 | VSArray x2 = vs_array_copy(x); | |
231 | double h = rand()%2 ? 1e-6 : -1e-6; | |
232 | x2.dat[k]+=h; | |
233 | double v2 = eval(x2, dat); | |
234 | vs_array_zero(&grad); | |
235 | grad.dat[k] = (v - v2)/h; | |
236 | vs_array_plus(&x2, x, *vs_array_scale(&x2, grad, stepsizes.dat[k])); | |
237 | v2 = eval(x2, dat); | |
238 | if(v2 < v){ | |
239 | //fprintf(stderr,"+"); | |
240 | vs_array_free(x); | |
241 | x = x2; | |
242 | v = v2; | |
243 | stepsizes.dat[k]*=1.2; // increase stepsize (4 successful steps will double it) | |
244 | }else{ // overshoot: reduce stepsize and don't do the step | |
245 | //fprintf(stderr,"."); | |
246 | stepsizes.dat[k]/=2.0; | |
247 | vs_array_free(x2); | |
248 | } | |
249 | //if(k==3) fprintf(stderr," "); | |
250 | } | |
251 | vs_array_free(grad); | |
252 | vs_array_free(stepsizes); | |
253 | if(residual != NULL) *residual=v; | |
254 | return x; | |
255 | } | |
256 | ||
257 | ||
258 | /* *** old calculation ***/ | |
259 | ||
260 | /* calculates rotation angle for the given transform and | |
261 | * field with respect to the given center-point | |
262 | */ | |
263 | double vsCalcAngle(const LocalMotion* lm, int center_x, int center_y){ | |
264 | // we better ignore fields that are to close to the rotation center | |
265 | if (abs(lm->f.x - center_x) + abs(lm->f.y - center_y) < lm->f.size*2) { | |
266 | return 0; | |
267 | } else { | |
268 | // double r = sqrt(lm->f.x*lm->f.x + lm->f.y*lm->f.y); | |
269 | double a1 = atan2(lm->f.y - center_y, lm->f.x - center_x); | |
270 | double a2 = atan2(lm->f.y - center_y + lm->v.y, | |
271 | lm->f.x - center_x + lm->v.x); | |
272 | double diff = a2 - a1; | |
273 | return (diff > M_PI) ? diff - 2 * M_PI : ((diff < -M_PI) ? diff + 2 | |
274 | * M_PI : diff); | |
275 | } | |
276 | } | |
277 | ||
278 | ||
279 | VSTransform vsSimpleMotionsToTransform(VSFrameInfo fi, const char* modName, | |
280 | const LocalMotions* motions){ | |
281 | int center_x = 0; | |
282 | int center_y = 0; | |
283 | VSTransform t = null_transform(); | |
284 | if(motions==0) return t; | |
285 | int num_motions=vs_vector_size(motions); | |
286 | double *angles = (double*) vs_malloc(sizeof(double) * num_motions); | |
287 | LocalMotion meanmotion; | |
288 | int i; | |
289 | if(num_motions < 1) | |
290 | return t; | |
291 | ||
292 | // calc center point of all remaining fields | |
293 | for (i = 0; i < num_motions; i++) { | |
294 | center_x += LMGet(motions,i)->f.x; | |
295 | center_y += LMGet(motions,i)->f.y; | |
296 | } | |
297 | center_x /= num_motions; | |
298 | center_y /= num_motions; | |
299 | ||
300 | // cleaned mean | |
301 | meanmotion = cleanmean_localmotions(motions); | |
302 | ||
303 | // figure out angle | |
304 | if (num_motions < 6) { | |
305 | // the angle calculation is inaccurate for 5 and less fields | |
306 | t.alpha = 0; | |
307 | } else { | |
308 | for (i = 0; i < num_motions; i++) { | |
309 | // substract avg and calc angle | |
310 | LocalMotion m = sub_localmotion(LMGet(motions,i),&meanmotion); | |
311 | angles[i] = vsCalcAngle(&m, center_x, center_y); | |
312 | } | |
313 | double min, max; | |
314 | t.alpha = -cleanmean(angles, num_motions, &min, &max); | |
315 | if (max - min > 1.0) { | |
316 | t.alpha = 0; | |
317 | vs_log_info(modName, "too large variation in angle(%f)\n", | |
318 | max-min); | |
319 | } | |
320 | } | |
321 | vs_free(angles); | |
322 | // compensate for off-center rotation | |
323 | double p_x = (center_x - fi.width / 2); | |
324 | double p_y = (center_y - fi.height / 2); | |
325 | t.x = meanmotion.v.x + (cos(t.alpha) - 1) * p_x - sin(t.alpha) * p_y; | |
326 | t.y = meanmotion.v.y + sin(t.alpha) * p_x + (cos(t.alpha) - 1) * p_y; | |
327 | ||
328 | return t; | |
329 | } | |
330 | ||
331 | ||
332 | ||
333 | /* | |
334 | * Local variables: | |
335 | * c-file-style: "stroustrup" | |
336 | * c-file-offsets: ((case-label . *) (statement-case-intro . *)) | |
337 | * indent-tabs-mode: nil | |
338 | * c-basic-offset: 2 t | |
339 | * End: | |
340 | * | |
341 | * vim: expandtab shiftwidth=2: | |
342 | */ |