Changeset 30305

Timestamp:

Jan 19, 2011, 3:23:09 PM (15 years ago)

Author:

eugene

Message:

back to using weighted fits so the chisq and penalty is generally sensible; adjust the chisq-based score using the sum-square of the smoothing kernel -- this automatically accounts for larger output images

File:

• branches/eam_branches/ipp-20101205/psModules/src/imcombine/pmSubtractionEquation.c (modified) (3 diffs)

branches/eam_branches/ipp-20101205/psModules/src/imcombine/pmSubtractionEquation.c

@@ -18 +18 @@
 
 //# define TESTING                         // TESTING output for debugging; may not work with threads!
-//# define USE_WEIGHT                      // Include weight (1/variance) in equation?
+# define USE_WEIGHT                      // Include weight (1/variance) in equation?
 # define USE_WINDOW                      // window to avoid neighbor contamination
 
@@ -1589 +1589 @@
     float momentValue = stats->sampleMean;
 
-    // XXX this is probably wrong : I want to score against higher moments **compared with the raw momements**
-    // score : chisq * dMoments * modeFactor
-    float modeFactor = kernels->mode == PM_SUBTRACTION_MODE_DUAL ? 2.0 : 1.0;
-    float orderFactor = 0.5*(kernels->spatialOrder + 1) * (kernels->spatialOrder + 2);
-    float score = chisqValue * momentValue * modeFactor * orderFactor;
-
-    fprintf (stderr, "chisq: %f, moment: %f, score: %f\n", chisqValue, momentValue, score);
+    double sumKernel1 = 0.0, sumKernel2 = 0.0; // Sum of the kernel
+
+    // calculate the variance contribution from this smoothing kernel
+    psKernel *modelKernel = pmSubtractionKernel(kernels, 0.0, 0.0, false);
+    for (int y = modelKernel->yMin; y <= modelKernel->yMax; y++) {
+        for (int x = modelKernel->xMin; x <= modelKernel->xMax; x++) {
+            if (!isfinite(modelKernel->kernel[y][x])) {
+                psError(PS_ERR_BAD_PARAMETER_VALUE, true, "Non-finite covariance matrix element in kernel at %d,%d", x, y);
+                return NULL;
+            }
+            sumKernel1 += PS_SQR(modelKernel->kernel[y][x]);
+        }
+    }
+    psFree (modelKernel);
+
+    if (kernels->mode == PM_SUBTRACTION_MODE_DUAL) {
+        psKernel *modelKernel = pmSubtractionKernel(kernels, 0.0, 0.0, true);
+        for (int y = modelKernel->yMin; y <= modelKernel->yMax; y++) {
+            for (int x = modelKernel->xMin; x <= modelKernel->xMax; x++) {
+                if (!isfinite(modelKernel->kernel[y][x])) {
+                    psError(PS_ERR_BAD_PARAMETER_VALUE, true, "Non-finite covariance matrix element in kernel at %d,%d", x, y);
+                    return NULL;
+                }
+                sumKernel2 += PS_SQR(modelKernel->kernel[y][x]);
+            }
+        }
+        psFree (modelKernel);
+    } else {
+        sumKernel2 = 1.0;
+    }
+
+    // if we modify the chisq value by the (sumKernel1 + sumKernel2), we account for the
+    // smoothing coming from larger kernels adding additional spatial fit terms should be
+    // penalized by increasing the score somewhat.  the 0.01 value is not well-chosen.
+    float orderFactor = 0.01 * kernels->spatialOrder;
+    float score = 2.0 * chisqValue / (sumKernel1 + sumKernel2) + orderFactor;
+    psLogMsg("psModules.imcombine", PS_LOG_INFO, "chisq: %6.3f, moment: %6.3f, sumKernel_1: %6.3f, sumKernel_2, score: %6.3f: %6.3f\n", chisqValue, momentValue, sumKernel1, sumKernel2, score);
 
     // save this result if it is the first or the best (skip if bestMatch is NULL)
@@ -1614 +1644 @@
         }
         if (keep) {
-            fprintf (stderr, "keeping order: %d, mode: %d, score: %f\n", kernels->spatialOrder, kernels->mode, score);
+            psLogMsg("psModules.imcombine", PS_LOG_INFO, "keeping order: %d, mode: %d, score: %f\n", kernels->spatialOrder, kernels->mode, score);
             match->score        = score;
             match->spatialOrder = kernels->spatialOrder;