• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (C) 2011 The Android Open Source Project
3  *
4  * Licensed under the Apache License, Version 2.0 (the "License");
5  * you may not use this file except in compliance with the License.
6  * You may obtain a copy of the License at
7  *
8  *      http://www.apache.org/licenses/LICENSE-2.0
9  *
10  * Unless required by applicable law or agreed to in writing, software
11  * distributed under the License is distributed on an "AS IS" BASIS,
12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13  * See the License for the specific language governing permissions and
14  * limitations under the License.
15  */
16 
17 /* $Id: db_metrics.h,v 1.3 2011/06/17 14:03:31 mbansal Exp $ */
18 
19 #ifndef DB_METRICS
20 #define DB_METRICS
21 
22 
23 
24 /*****************************************************************
25 *    Lean and mean begins here                                   *
26 *****************************************************************/
27 
28 #include "db_utilities.h"
29 /*!
30  * \defgroup LMMetrics (LM) Metrics
31  */
32 /*\{*/
33 
34 
35 
36 
37 /*!
38 Compute function value fp and Jacobian J of robustifier given input value f*/
db_CauchyDerivative(double J[4],double fp[2],const double f[2],double one_over_scale2)39 inline void db_CauchyDerivative(double J[4],double fp[2],const double f[2],double one_over_scale2)
40 {
41     double x2,y2,r,r2,r2s,one_over_r2,fu,r_fu,one_over_r_fu;
42     double one_plus_r2s,half_dfu_dx,half_dfu_dy,coeff,coeff2,coeff3;
43     int at_zero;
44 
45     /*The robustifier takes the input (x,y) and makes a new
46     vector (xp,yp) where
47     xp=sqrt(log(1+(x^2+y^2)*one_over_scale2))*x/sqrt(x^2+y^2)
48     yp=sqrt(log(1+(x^2+y^2)*one_over_scale2))*y/sqrt(x^2+y^2)
49     The new vector has the property
50     xp^2+yp^2=log(1+(x^2+y^2)*one_over_scale2)
51     i.e. when it is square-summed it gives the robust
52     reprojection error
53     Define
54     r2=(x^2+y^2) and
55     r2s=r2*one_over_scale2
56     fu=log(1+r2s)/r2
57     then
58     xp=sqrt(fu)*x
59     yp=sqrt(fu)*y
60     and
61     d(r2)/dx=2x
62     d(r2)/dy=2y
63     and
64     dfu/dx=d(r2)/dx*(r2s/(1+r2s)-log(1+r2s))/(r2*r2)
65     dfu/dy=d(r2)/dy*(r2s/(1+r2s)-log(1+r2s))/(r2*r2)
66     and
67     d(xp)/dx=1/(2sqrt(fu))*(dfu/dx)*x+sqrt(fu)
68     d(xp)/dy=1/(2sqrt(fu))*(dfu/dy)*x
69     d(yp)/dx=1/(2sqrt(fu))*(dfu/dx)*y
70     d(yp)/dy=1/(2sqrt(fu))*(dfu/dy)*y+sqrt(fu)
71     */
72 
73     x2=db_sqr(f[0]);
74     y2=db_sqr(f[1]);
75     r2=x2+y2;
76     r=sqrt(r2);
77 
78     if(r2<=0.0) at_zero=1;
79     else
80     {
81         one_over_r2=1.0/r2;
82         r2s=r2*one_over_scale2;
83         one_plus_r2s=1.0+r2s;
84         fu=log(one_plus_r2s)*one_over_r2;
85         r_fu=sqrt(fu);
86         if(r_fu<=0.0) at_zero=1;
87         else
88         {
89             one_over_r_fu=1.0/r_fu;
90             fp[0]=r_fu*f[0];
91             fp[1]=r_fu*f[1];
92             /*r2s is always >= 0*/
93             coeff=(r2s/one_plus_r2s*one_over_r2-fu)*one_over_r2;
94             half_dfu_dx=f[0]*coeff;
95             half_dfu_dy=f[1]*coeff;
96             coeff2=one_over_r_fu*half_dfu_dx;
97             coeff3=one_over_r_fu*half_dfu_dy;
98 
99             J[0]=coeff2*f[0]+r_fu;
100             J[1]=coeff3*f[0];
101             J[2]=coeff2*f[1];
102             J[3]=coeff3*f[1]+r_fu;
103             at_zero=0;
104         }
105     }
106     if(at_zero)
107     {
108         /*Close to zero the robustifying mapping
109         becomes identity*sqrt(one_over_scale2)*/
110         fp[0]=0.0;
111         fp[1]=0.0;
112         J[0]=sqrt(one_over_scale2);
113         J[1]=0.0;
114         J[2]=0.0;
115         J[3]=J[0];
116     }
117 }
118 
db_SquaredReprojectionErrorHomography(const double y[2],const double H[9],const double x[3])119 inline double db_SquaredReprojectionErrorHomography(const double y[2],const double H[9],const double x[3])
120 {
121     double x0,x1,x2,mult;
122     double sd;
123 
124     x0=H[0]*x[0]+H[1]*x[1]+H[2]*x[2];
125     x1=H[3]*x[0]+H[4]*x[1]+H[5]*x[2];
126     x2=H[6]*x[0]+H[7]*x[1]+H[8]*x[2];
127     mult=1.0/((x2!=0.0)?x2:1.0);
128     sd=db_sqr((y[0]-x0*mult))+db_sqr((y[1]-x1*mult));
129 
130     return(sd);
131 }
132 
db_SquaredInhomogenousHomographyError(const double y[2],const double H[9],const double x[2])133 inline double db_SquaredInhomogenousHomographyError(const double y[2],const double H[9],const double x[2])
134 {
135     double x0,x1,x2,mult;
136     double sd;
137 
138     x0=H[0]*x[0]+H[1]*x[1]+H[2];
139     x1=H[3]*x[0]+H[4]*x[1]+H[5];
140     x2=H[6]*x[0]+H[7]*x[1]+H[8];
141     mult=1.0/((x2!=0.0)?x2:1.0);
142     sd=db_sqr((y[0]-x0*mult))+db_sqr((y[1]-x1*mult));
143 
144     return(sd);
145 }
146 
147 /*!
148 Return a constant divided by likelihood of a Cauchy distributed
149 reprojection error given the image point y, homography H, image point
150 point x and the squared scale coefficient one_over_scale2=1.0/(scale*scale)
151 where scale is the half width at half maximum (hWahM) of the
152 Cauchy distribution*/
db_ExpCauchyInhomogenousHomographyError(const double y[2],const double H[9],const double x[2],double one_over_scale2)153 inline double db_ExpCauchyInhomogenousHomographyError(const double y[2],const double H[9],const double x[2],
154                                                       double one_over_scale2)
155 {
156     double sd;
157     sd=db_SquaredInhomogenousHomographyError(y,H,x);
158     return(1.0+sd*one_over_scale2);
159 }
160 
161 /*!
162 Compute residual vector f between image point y and homography Hx of
163 image point x. Also compute Jacobian of f with respect
164 to an update dx of H*/
db_DerivativeInhomHomographyError(double Jf_dx[18],double f[2],const double y[2],const double H[9],const double x[2])165 inline void db_DerivativeInhomHomographyError(double Jf_dx[18],double f[2],const double y[2],const double H[9],
166                                               const double x[2])
167 {
168     double xh,yh,zh,mult,mult2,xh_mult2,yh_mult2;
169     /*The Jacobian of the inhomogenous coordinates with respect to
170     the homogenous is
171     [1/zh  0  -xh/(zh*zh)]
172     [ 0  1/zh -yh/(zh*zh)]
173     The Jacobian of the homogenous coordinates with respect to dH is
174     [x0 x1 1  0  0 0  0  0 0]
175     [ 0  0 0 x0 x1 1  0  0 0]
176     [ 0  0 0  0  0 0 x0 x1 1]
177     The output Jacobian is minus their product, i.e.
178     [-x0/zh -x1/zh -1/zh    0      0     0    x0*xh/(zh*zh) x1*xh/(zh*zh) xh/(zh*zh)]
179     [   0      0     0   -x0/zh -x1/zh -1/zh  x0*yh/(zh*zh) x1*yh/(zh*zh) yh/(zh*zh)]*/
180 
181     /*Compute warped point, which is the same as
182     homogenous coordinates of reprojection*/
183     xh=H[0]*x[0]+H[1]*x[1]+H[2];
184     yh=H[3]*x[0]+H[4]*x[1]+H[5];
185     zh=H[6]*x[0]+H[7]*x[1]+H[8];
186     mult=1.0/((zh!=0.0)?zh:1.0);
187     /*Compute inhomogenous residual*/
188     f[0]=y[0]-xh*mult;
189     f[1]=y[1]-yh*mult;
190     /*Compute Jacobian*/
191     mult2=mult*mult;
192     xh_mult2=xh*mult2;
193     yh_mult2=yh*mult2;
194     Jf_dx[0]= -x[0]*mult;
195     Jf_dx[1]= -x[1]*mult;
196     Jf_dx[2]= -mult;
197     Jf_dx[3]=0;
198     Jf_dx[4]=0;
199     Jf_dx[5]=0;
200     Jf_dx[6]=x[0]*xh_mult2;
201     Jf_dx[7]=x[1]*xh_mult2;
202     Jf_dx[8]=xh_mult2;
203     Jf_dx[9]=0;
204     Jf_dx[10]=0;
205     Jf_dx[11]=0;
206     Jf_dx[12]=Jf_dx[0];
207     Jf_dx[13]=Jf_dx[1];
208     Jf_dx[14]=Jf_dx[2];
209     Jf_dx[15]=x[0]*yh_mult2;
210     Jf_dx[16]=x[1]*yh_mult2;
211     Jf_dx[17]=yh_mult2;
212 }
213 
214 /*!
215 Compute robust residual vector f between image point y and homography Hx of
216 image point x. Also compute Jacobian of f with respect
217 to an update dH of H*/
db_DerivativeCauchyInhomHomographyReprojection(double Jf_dx[18],double f[2],const double y[2],const double H[9],const double x[2],double one_over_scale2)218 inline void db_DerivativeCauchyInhomHomographyReprojection(double Jf_dx[18],double f[2],const double y[2],const double H[9],
219                                                            const double x[2],double one_over_scale2)
220 {
221     double Jf_dx_loc[18],f_loc[2];
222     double J[4],J0,J1,J2,J3;
223 
224     /*Compute reprojection Jacobian*/
225     db_DerivativeInhomHomographyError(Jf_dx_loc,f_loc,y,H,x);
226     /*Compute robustifier Jacobian*/
227     db_CauchyDerivative(J,f,f_loc,one_over_scale2);
228 
229     /*Multiply the robustifier Jacobian with
230     the reprojection Jacobian*/
231     J0=J[0];J1=J[1];J2=J[2];J3=J[3];
232     Jf_dx[0]=J0*Jf_dx_loc[0];
233     Jf_dx[1]=J0*Jf_dx_loc[1];
234     Jf_dx[2]=J0*Jf_dx_loc[2];
235     Jf_dx[3]=                J1*Jf_dx_loc[12];
236     Jf_dx[4]=                J1*Jf_dx_loc[13];
237     Jf_dx[5]=                J1*Jf_dx_loc[14];
238     Jf_dx[6]=J0*Jf_dx_loc[6]+J1*Jf_dx_loc[15];
239     Jf_dx[7]=J0*Jf_dx_loc[7]+J1*Jf_dx_loc[16];
240     Jf_dx[8]=J0*Jf_dx_loc[8]+J1*Jf_dx_loc[17];
241     Jf_dx[9]= J2*Jf_dx_loc[0];
242     Jf_dx[10]=J2*Jf_dx_loc[1];
243     Jf_dx[11]=J2*Jf_dx_loc[2];
244     Jf_dx[12]=                J3*Jf_dx_loc[12];
245     Jf_dx[13]=                J3*Jf_dx_loc[13];
246     Jf_dx[14]=                J3*Jf_dx_loc[14];
247     Jf_dx[15]=J2*Jf_dx_loc[6]+J3*Jf_dx_loc[15];
248     Jf_dx[16]=J2*Jf_dx_loc[7]+J3*Jf_dx_loc[16];
249     Jf_dx[17]=J2*Jf_dx_loc[8]+J3*Jf_dx_loc[17];
250 }
251 /*!
252 Compute residual vector f between image point y and rotation of
253 image point x by R. Also compute Jacobian of f with respect
254 to an update dx of R*/
db_DerivativeInhomRotationReprojection(double Jf_dx[6],double f[2],const double y[2],const double R[9],const double x[2])255 inline void db_DerivativeInhomRotationReprojection(double Jf_dx[6],double f[2],const double y[2],const double R[9],
256                                                    const double x[2])
257 {
258     double xh,yh,zh,mult,mult2,xh_mult2,yh_mult2;
259     /*The Jacobian of the inhomogenous coordinates with respect to
260     the homogenous is
261     [1/zh  0  -xh/(zh*zh)]
262     [ 0  1/zh -yh/(zh*zh)]
263     The Jacobian at zero of the homogenous coordinates with respect to
264     [sin(phi) sin(ohm) sin(kap)] is
265     [-rx2   0   rx1 ]
266     [  0   rx2 -rx0 ]
267     [ rx0 -rx1   0  ]
268     The output Jacobian is minus their product, i.e.
269     [1+xh*xh/(zh*zh) -xh*yh/(zh*zh)   -yh/zh]
270     [xh*yh/(zh*zh)   -1-yh*yh/(zh*zh)  xh/zh]*/
271 
272     /*Compute rotated point, which is the same as
273     homogenous coordinates of reprojection*/
274     xh=R[0]*x[0]+R[1]*x[1]+R[2];
275     yh=R[3]*x[0]+R[4]*x[1]+R[5];
276     zh=R[6]*x[0]+R[7]*x[1]+R[8];
277     mult=1.0/((zh!=0.0)?zh:1.0);
278     /*Compute inhomogenous residual*/
279     f[0]=y[0]-xh*mult;
280     f[1]=y[1]-yh*mult;
281     /*Compute Jacobian*/
282     mult2=mult*mult;
283     xh_mult2=xh*mult2;
284     yh_mult2=yh*mult2;
285     Jf_dx[0]= 1.0+xh*xh_mult2;
286     Jf_dx[1]= -yh*xh_mult2;
287     Jf_dx[2]= -yh*mult;
288     Jf_dx[3]= -Jf_dx[1];
289     Jf_dx[4]= -1-yh*yh_mult2;
290     Jf_dx[5]= xh*mult;
291 }
292 
293 /*!
294 Compute robust residual vector f between image point y and rotation of
295 image point x by R. Also compute Jacobian of f with respect
296 to an update dx of R*/
db_DerivativeCauchyInhomRotationReprojection(double Jf_dx[6],double f[2],const double y[2],const double R[9],const double x[2],double one_over_scale2)297 inline void db_DerivativeCauchyInhomRotationReprojection(double Jf_dx[6],double f[2],const double y[2],const double R[9],
298                                                          const double x[2],double one_over_scale2)
299 {
300     double Jf_dx_loc[6],f_loc[2];
301     double J[4],J0,J1,J2,J3;
302 
303     /*Compute reprojection Jacobian*/
304     db_DerivativeInhomRotationReprojection(Jf_dx_loc,f_loc,y,R,x);
305     /*Compute robustifier Jacobian*/
306     db_CauchyDerivative(J,f,f_loc,one_over_scale2);
307 
308     /*Multiply the robustifier Jacobian with
309     the reprojection Jacobian*/
310     J0=J[0];J1=J[1];J2=J[2];J3=J[3];
311     Jf_dx[0]=J0*Jf_dx_loc[0]+J1*Jf_dx_loc[3];
312     Jf_dx[1]=J0*Jf_dx_loc[1]+J1*Jf_dx_loc[4];
313     Jf_dx[2]=J0*Jf_dx_loc[2]+J1*Jf_dx_loc[5];
314     Jf_dx[3]=J2*Jf_dx_loc[0]+J3*Jf_dx_loc[3];
315     Jf_dx[4]=J2*Jf_dx_loc[1]+J3*Jf_dx_loc[4];
316     Jf_dx[5]=J2*Jf_dx_loc[2]+J3*Jf_dx_loc[5];
317 }
318 
319 
320 
321 /*!
322 // remove the outliers whose projection error is larger than pre-defined
323 */
324 inline int db_RemoveOutliers_Homography(const double H[9], double *x_i,double *xp_i, double *wp,double *im, double *im_p, double *im_r, double *im_raw,double *im_raw_p,int point_count,double scale, double thresh=DB_OUTLIER_THRESHOLD)
325 {
326     double temp_valueE, t2;
327     int c;
328     int k1=0;
329     int k2=0;
330     int k3=0;
331     int numinliers=0;
332     int ind1;
333     int ind2;
334     int ind3;
335     int isinlier;
336 
337     // experimentally determined
338     t2=1.0/(thresh*thresh*thresh*thresh);
339 
340     // count the inliers
341     for(c=0;c<point_count;c++)
342     {
343         ind1=c<<1;
344         ind2=c<<2;
345         ind3=3*c;
346 
347         temp_valueE=db_SquaredInhomogenousHomographyError(im_p+ind3,H,im+ind3);
348 
349         isinlier=((temp_valueE<=t2)?1:0);
350 
351         // if it is inlier, then copy the 3d and 2d correspondences
352         if (isinlier)
353         {
354             numinliers++;
355 
356             x_i[k1]=x_i[ind1];
357             x_i[k1+1]=x_i[ind1+1];
358 
359             xp_i[k1]=xp_i[ind1];
360             xp_i[k1+1]=xp_i[ind1+1];
361 
362             k1=k1+2;
363 
364             // original normalized pixel coordinates
365             im[k3]=im[ind3];
366             im[k3+1]=im[ind3+1];
367             im[k3+2]=im[ind3+2];
368 
369             im_r[k3]=im_r[ind3];
370             im_r[k3+1]=im_r[ind3+1];
371             im_r[k3+2]=im_r[ind3+2];
372 
373             im_p[k3]=im_p[ind3];
374             im_p[k3+1]=im_p[ind3+1];
375             im_p[k3+2]=im_p[ind3+2];
376 
377             // left and right raw pixel coordinates
378             im_raw[k3] = im_raw[ind3];
379             im_raw[k3+1] = im_raw[ind3+1];
380             im_raw[k3+2] = im_raw[ind3+2]; // the index
381 
382             im_raw_p[k3] = im_raw_p[ind3];
383             im_raw_p[k3+1] = im_raw_p[ind3+1];
384             im_raw_p[k3+2] = im_raw_p[ind3+2]; // the index
385 
386             k3=k3+3;
387 
388             // 3D coordinates
389             wp[k2]=wp[ind2];
390             wp[k2+1]=wp[ind2+1];
391             wp[k2+2]=wp[ind2+2];
392             wp[k2+3]=wp[ind2+3];
393 
394             k2=k2+4;
395 
396         }
397     }
398 
399     return numinliers;
400 }
401 
402 
403 
404 
405 
406 /*\}*/
407 
408 #endif /* DB_METRICS */
409