localmotion2transform.c

/*
 * localmotion2transform.c
 *
 *  Copyright (C) Georg Martius - January 2013
 *   georg dot martius at web dot de
 *
 *  This file is part of vid.stab video stabilization library
 *
 *  vid.stab is free software; you can redistribute it and/or modify
 *  it under the terms of the GNU General Public License,
 *  as published by the Free Software Foundation; either version 2, or
 *  (at your option) any later version.
 *
 *  vid.stab is distributed in the hope that it will be useful,
 *  but WITHOUT ANY WARRANTY; without even the implied warranty of
 *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 *  GNU General Public License for more details.
 *
 *  You should have received a copy of the GNU General Public License
 *  along with GNU Make; see the file COPYING.  If not, write to
 *  the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
 *
 */

#include "localmotion2transform.h"
#include "transformtype_operations.h"
#include <assert.h>
#include <string.h>

/* #include <sys/time.h> */
/* long timeOfDayinMS() { */
/*   struct timeval t; */
/*   gettimeofday(&t, 0); */
/*   return t.tv_sec*1000 + t.tv_usec/1000; */
/* } */

int vsLocalmotions2Transforms(VSTransformData* td,
                              const VSManyLocalMotions* motions,
                              VSTransformations* trans ){
  int len = vs_vector_size(motions);
  assert(trans->len==0 && trans->ts == 0);
  trans->ts = vs_malloc(sizeof(VSTransform)*len );
  /* long start= timeOfDayinMS(); */
  FILE *f=0;
  if(td->conf.storeTransforms){
    f = fopen("global_motions.trf","w");
  }

  if(td->conf.simpleMotionCalculation==0){
    for(int i=0; i< vs_vector_size(motions); i++) {
      trans->ts[i]=vsMotionsToTransform(td,VSMLMGet(motions,i), f);
    }
  }else{
    for(int i=0; i< vs_vector_size(motions); i++) {
      trans->ts[i]=vsSimpleMotionsToTransform(td->fiSrc, td->conf.modName,VSMLMGet(motions,i));
    }
  }
  trans->len=len;

  /* long end = timeOfDayinMS(); */
  /* vs_log_info(td->conf.modName, "Localmotions2Transform (%i) with %i frames took %i ms\n", */
  /*             td->conf.simpleMotionCalculation, len, end-start); */
  if(f) fclose(f);
  return VS_OK;
}

VSArray vsTransformToArray(const VSTransform* t){
  VSArray a = vs_array_new(4);
  a.dat[0] = t->x;
  a.dat[1] = t->y;
  a.dat[2] = t->alpha;
  a.dat[3] = t->zoom;
  return a;
}

VSTransform vsArrayToTransform(VSArray a){
  return new_transform(a.dat[0],a.dat[1],a.dat[2],a.dat[3],0,0,0);
}

struct VSGradientDat {
  VSTransformData* td;
  const LocalMotions* motions;
  VSArray missmatches; // if negative then local motion is ignored
};

double calcTransformQuality(VSArray params, void* dat){
  struct VSGradientDat* gd= (struct VSGradientDat*) dat;
  const LocalMotions* motions = gd->motions;
  int num_motions=vs_vector_size(motions);
  VSTransform t = vsArrayToTransform(params);
  double error=0;

  PreparedTransform pt= prepare_transform(&t, &gd->td->fiSrc);
  int num = 1; // we start with 1 to avoid div by zero
  for (int i = 0; i < num_motions; i++) {
    if(gd->missmatches.dat[i]>=0){
      LocalMotion* m = LMGet(motions,i);
      double vx,vy;
      transform_vec_double(&vx, &vy, &pt, (Vec*)&m->f);
      vx -= m->f.x; vy -= m->f.y;
      double e   = sqr(vx - m->v.x) +  sqr(vy - m->v.y);
      gd->missmatches.dat[i]=e;
      error += e;
      num++;
    }
  }
  // 1 pixel translation missmatch is roughly (with size 500):
  // alpha=0.11 (degree), zoom=0.2; The zoom is however often much larger, so less penalty.
  return error/num + fabs(t.alpha)/5.0 + fabs(t.zoom)/500.0;
}

double intMean(const int* ds, int len) {
  double sum=0;
  for (int i = 0; i < len; i++) sum += ds[i];
  return sum / len;
}

// only calcates means transform to initialise gradient descent
VSTransform meanMotions(VSTransformData* td, const LocalMotions* motions){
  int len = vs_vector_size(motions);
  int* xs = localmotions_getx(motions);
  int* ys = localmotions_gety(motions);
  VSTransform t = null_transform();
  if(motions==0 || len==0) {
    t.extra = 1; // prob. blank frame or too low contrast, ignore later
    return t;
  }
  t.x = intMean(xs,len);
  t.y = intMean(ys,len);
  vs_free(xs);
  vs_free(ys);
  return t;
}

/* Disables those fields (mask = -1) whose (miss)quality is high.
   @param mask: fields masks (<0 means disabled)
   @param missqualities: measure for each field (larger is worse)
   @param stddevs: x standard deviations to exclude
   Both array have to be of the same length.
   @return number of disabled fields
*/
int disableFields(VSArray mask, VSArray missqualities, double stddevs){
  assert(mask.len == missqualities.len);
  // first we throw away those fields that match badely (during motion detection)
  double mu   = mean(missqualities.dat, missqualities.len);
  double sigma = stddev(missqualities.dat, missqualities.len, mu);
  double thresh = mu + stddevs * sigma;
  int cnt=0;
  for(int i=0; i< mask.len; i++){
    if(missqualities.dat[i]>thresh){
      mask.dat[i]=-1.0; // disable field
      cnt++;
    }
  }
  return cnt;
}

VSTransform vsMotionsToTransform(VSTransformData* td,
                                 const LocalMotions* motions,
                                 FILE* f){
  VSTransform t = meanMotions(td, motions);
  if(motions==0 || vs_vector_size(motions)==0){
    if (f) fprintf(f,"0 0 0 0 0 %i\n# no fields\n", t.extra);
    return t;
  }
  VSArray missmatches = vs_array_new(vs_vector_size(motions));
  VSArray params = vsTransformToArray(&t);
  double residual;
  struct VSGradientDat dat;
  dat.motions = motions;
  dat.td      = td;
  dat.missmatches = missmatches;

  // first we throw away those fields that match badely (during motion detection)
  VSArray matchQualities = localmotionsGetMatch(motions);
  int dis1=disableFields(missmatches, matchQualities, 1.5);
  vs_array_free(matchQualities);

  VSArray result;
  double ss[] = {0.2, 0.2, 0.00005, 0.1};
  int k;
  int dis2=0;
  for(k=0; k<3; k++){
    // optimize params to minimize transform quality (12 steps per dimension)
    result = vsGradientDescent(calcTransformQuality, params, &dat,
                                        16, vs_array(ss,4), 0.01, &residual);
    vs_array_free(params);
    // now we need to ignore the fields that don't fit well (e.g. moving objects)
    // cut off everthing above 1 std. dev. for skewed distributions
    // this will cut off the tail
    // do this only two times (3 gradient optimizations in total)
    if((k==0 && residual>0.1) || (k==1 && residual>20)){
      dis2 += disableFields(missmatches, missmatches, 1.0);
      params = result;
    } else break;
  }

  if(td->conf.verbose  & VS_DEBUG)
    vs_log_info(td->conf.modName, "disabled (%i+%i)/%i,\tresidual: %f (%i)\n",
                dis1, dis2, vs_vector_size(motions), residual, k+1);
  t = vsArrayToTransform(result);
  vs_array_free(result);
  vs_array_free(missmatches);
  // check if sufficiently good match was achieved:
  if(residual>100){ // test threshold.
    t.extra=1;
  }
  if(f){
    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,
            residual, k + 1);
  }
  if(!td->conf.smoothZoom)
    t.zoom=0;
  return t;
}



/* n-dimensional general purpose gradient descent algorithm */
VSArray vsGradientDescent(double (*eval)(VSArray, void*),
                         VSArray params, void* dat,
                         int N, VSArray stepsizes, double threshold, double* residual){
  int dim=params.len;
  double v = eval(params, dat);
  VSArray x = vs_array_copy(params);
  VSArray grad = vs_array_new(dim);
  assert(stepsizes.len == params.len);
  for(int i=0; i< N*dim && v > threshold; i++){
    int k=i%dim;
    VSArray x2 = vs_array_copy(x);
    double h = rand()%2 ? 1e-6 : -1e-6;
    x2.dat[k]+=h;
    double v2 = eval(x2, dat);
    vs_array_zero(&grad);
    grad.dat[k] = (v - v2)/h;
    vs_array_plus(&x2, x, *vs_array_scale(&x2, grad, stepsizes.dat[k]));
    v2 = eval(x2, dat);
    if(v2 < v){
      //fprintf(stderr,"+");
      vs_array_free(x);
      x = x2;
      v = v2;
      stepsizes.dat[k]*=1.2; // increase stepsize (4 successful steps will double it)
    }else{ // overshoot: reduce stepsize and don't do the step
      //fprintf(stderr,".");
      stepsizes.dat[k]/=2.0;
      vs_array_free(x2);
    }
    //if(k==3) fprintf(stderr," ");
  }
  vs_array_free(grad);
  vs_array_free(stepsizes);
  if(residual != NULL) *residual=v;
  return x;
}


/* *** old  calculation ***/

/* calculates rotation angle for the given transform and
 * field with respect to the given center-point
 */
double vsCalcAngle(const LocalMotion* lm, int center_x, int center_y){
  // we better ignore fields that are to close to the rotation center
  if (abs(lm->f.x - center_x) + abs(lm->f.y - center_y) < lm->f.size*2) {
    return 0;
  } else {
    // double r = sqrt(lm->f.x*lm->f.x + lm->f.y*lm->f.y);
    double a1 = atan2(lm->f.y - center_y, lm->f.x - center_x);
    double a2 = atan2(lm->f.y - center_y + lm->v.y,
                      lm->f.x - center_x + lm->v.x);
    double diff = a2 - a1;
    return (diff > M_PI) ? diff - 2 * M_PI : ((diff < -M_PI) ? diff + 2
                * M_PI : diff);
  }
}


VSTransform vsSimpleMotionsToTransform(VSFrameInfo fi, const char* modName,
                                       const LocalMotions* motions){
  int center_x = 0;
  int center_y = 0;
  VSTransform t = null_transform();
  if(motions==0) return t;
  int num_motions=vs_vector_size(motions);
  double *angles = (double*) vs_malloc(sizeof(double) * num_motions);
  LocalMotion meanmotion;
  int i;
  if(num_motions < 1)
    return t;

  // calc center point of all remaining fields
  for (i = 0; i < num_motions; i++) {
    center_x += LMGet(motions,i)->f.x;
    center_y += LMGet(motions,i)->f.y;
  }
  center_x /= num_motions;
  center_y /= num_motions;

  // cleaned mean
  meanmotion = cleanmean_localmotions(motions);

  // figure out angle
  if (num_motions < 6) {
    // the angle calculation is inaccurate for 5 and less fields
    t.alpha = 0;
  } else {
    for (i = 0; i < num_motions; i++) {
      // substract avg and calc angle
      LocalMotion m = sub_localmotion(LMGet(motions,i),&meanmotion);
      angles[i] = vsCalcAngle(&m, center_x, center_y);
    }
    double min, max;
    t.alpha = -cleanmean(angles, num_motions, &min, &max);
    if (max - min > 1.0) {
      t.alpha = 0;
      vs_log_info(modName, "too large variation in angle(%f)\n",
      max-min);
    }
  }
  vs_free(angles);
  // compensate for off-center rotation
  double p_x = (center_x - fi.width / 2);
  double p_y = (center_y - fi.height / 2);
  t.x = meanmotion.v.x + (cos(t.alpha) - 1) * p_x - sin(t.alpha) * p_y;
  t.y = meanmotion.v.y + sin(t.alpha) * p_x + (cos(t.alpha) - 1) * p_y;

  return t;
}



/*
 * Local variables:
 *   c-file-style: "stroustrup"
 *   c-file-offsets: ((case-label . *) (statement-case-intro . *))
 *   indent-tabs-mode: nil
 *   c-basic-offset: 2 t
 * End:
 *
 * vim: expandtab shiftwidth=2:
 */