Commit bdf3d47f authored by Andrey Filippov's avatar Andrey Filippov

splitting planes into pairs

parent 41d614cb
......@@ -2162,6 +2162,10 @@ public class EyesisCorrectionParameters {
public int plIterations = 10; // Maximal number of smoothing iterations for each step
public int plPrecision = 6; // Maximal step difference (1/power of 10)
public double plSplitPull = .5; // Relative weight of center plane when splitting into pairs
public int plSplitMinNeib = 2; // Minimal number of neighbors to split plane in pairs
public boolean plFuse = true; // Fuse planes together (off for debug only)
public boolean plKeepOrphans = true; // Keep unconnected supertiles
public double plMinOrphan = 2.0; // Minimal strength unconnected supertiles to keep
......@@ -2427,7 +2431,11 @@ public class EyesisCorrectionParameters {
properties.setProperty(prefix+"plPull", this.plPull +"");
properties.setProperty(prefix+"plIterations", this.plIterations+"");
properties.setProperty(prefix+"plPrecision", this.plPrecision+"");
properties.setProperty(prefix+"plFuse", this.plFuse+"");
properties.setProperty(prefix+"plSplitPull", this.plSplitPull +"");
properties.setProperty(prefix+"plSplitMinNeib", this.plSplitMinNeib+"");
properties.setProperty(prefix+"plFuse", this.plFuse+"");
properties.setProperty(prefix+"plKeepOrphans", this.plKeepOrphans+"");
properties.setProperty(prefix+"plMinOrphan", this.plMinOrphan +"");
......@@ -2680,6 +2688,10 @@ public class EyesisCorrectionParameters {
if (properties.getProperty(prefix+"plPull")!=null) this.plPull=Double.parseDouble(properties.getProperty(prefix+"plPull"));
if (properties.getProperty(prefix+"plIterations")!=null) this.plIterations=Integer.parseInt(properties.getProperty(prefix+"plIterations"));
if (properties.getProperty(prefix+"plPrecision")!=null) this.plPrecision=Integer.parseInt(properties.getProperty(prefix+"plPrecision"));
if (properties.getProperty(prefix+"plSplitPull")!=null) this.plSplitPull=Double.parseDouble(properties.getProperty(prefix+"plSplitPull"));
if (properties.getProperty(prefix+"plSplitMinNeib")!=null) this.plSplitMinNeib=Integer.parseInt(properties.getProperty(prefix+"plSplitMinNeib"));
if (properties.getProperty(prefix+"plFuse")!=null) this.plFuse=Boolean.parseBoolean(properties.getProperty(prefix+"plFuse"));
if (properties.getProperty(prefix+"plKeepOrphans")!=null) this.plKeepOrphans=Boolean.parseBoolean(properties.getProperty(prefix+"plKeepOrphans"));
if (properties.getProperty(prefix+"plMinOrphan")!=null) this.plMinOrphan=Double.parseDouble(properties.getProperty(prefix+"plMinOrphan"));
......@@ -2689,7 +2701,7 @@ public class EyesisCorrectionParameters {
if (properties.getProperty(prefix+"plSnapNegAny")!=null) this.plSnapNegAny=Double.parseDouble(properties.getProperty(prefix+"plSnapNegAny"));
if (properties.getProperty(prefix+"plSnapDispMax")!=null) this.plSnapDispMax=Double.parseDouble(properties.getProperty(prefix+"plSnapDispMax"));
if (properties.getProperty(prefix+"plSnapDispWeight")!=null) this.plSnapDispWeight=Double.parseDouble(properties.getProperty(prefix+"plSnapDispWeight"));
if (properties.getProperty(prefix+"plSnapZeroMode")!=null) this.plPrecision=Integer.parseInt(properties.getProperty(prefix+"plSnapZeroMode"));
if (properties.getProperty(prefix+"plSnapZeroMode")!=null) this.plPrecision=Integer.parseInt(properties.getProperty(prefix+"plSnapZeroMode"));
if (properties.getProperty(prefix+"show_ortho_combine")!=null) this.show_ortho_combine=Boolean.parseBoolean(properties.getProperty(prefix+"show_ortho_combine"));
if (properties.getProperty(prefix+"show_refine_supertiles")!=null) this.show_refine_supertiles=Boolean.parseBoolean(properties.getProperty(prefix+"show_refine_supertiles"));
......@@ -2955,6 +2967,10 @@ public class EyesisCorrectionParameters {
gd.addNumericField("Relative weight of original (measured) plane when combing with neighbors", this.plPull, 6);
gd.addNumericField("Maximal number of smoothing iterations for each step", this.plIterations, 0);
gd.addNumericField("Maximal step difference (1/power of 10)", this.plPrecision, 0);
gd.addNumericField("Relative weight of center plane when splitting into pairs", this.plSplitPull, 6);
gd.addNumericField("Minimal number of neighbors to split plane in pairs", this.plSplitMinNeib, 0);
gd.addCheckbox ("Fuse planes together (off for debug only)", this.plFuse);
gd.addCheckbox ("Keep unconnected supertiles", this.plKeepOrphans);
gd.addNumericField("Minimal strength unconnected supertiles to keep", this.plMinOrphan, 6);
......@@ -3218,6 +3234,10 @@ public class EyesisCorrectionParameters {
this.plPull= gd.getNextNumber();
this.plIterations= (int) gd.getNextNumber();
this.plPrecision= (int) gd.getNextNumber();
this.plSplitPull= gd.getNextNumber();
this.plSplitMinNeib= (int) gd.getNextNumber();
this.plFuse= gd.getNextBoolean();
this.plKeepOrphans= gd.getNextBoolean();
this.plMinOrphan= gd.getNextNumber();
......
......@@ -2680,6 +2680,205 @@ public class SuperTiles{
return diff; // return maximal difference
}
public double [][] planesGetDiff(
final TilePlanes.PlaneData[][] measured_planes,
final TilePlanes.PlaneData[][] mod_planes,
final int debugLevel,
final int dbg_X,
final int dbg_Y)
{
final int tilesX = tileProcessor.getTilesX();
final int tilesY = tileProcessor.getTilesY();
final int superTileSize = tileProcessor.getSuperTileSize();
final int stilesX = (tilesX + superTileSize -1)/superTileSize;
final int stilesY = (tilesY + superTileSize -1)/superTileSize;
final int debug_stile = dbg_Y * stilesX + dbg_X;
final TilePlanes.PlaneData[][] new_planes = copyPlanes(mod_planes);
final Thread[] threads = ImageDtt.newThreadArray(tileProcessor.threadsMax);
final double [][] diffs = null;
return diffs;
}
/**
* Create candidate planes to break a single plane in 2 by splitting consecutive connected
* neighbors in 2 groups that make the smallest weighted sum of the eigenvalues
* Groups may be optionally supplemented by the center supertile
* @param center_planes [per supertile][per plane] array of plane objects to use as
* the source of connections and optionally other data for the center tiles
* @param neib_planes [per supertile][per plane] array of plane objects to use for
* neighbor data. May be the same as center_planes or different
* @param center_pull - merge with center plane when calculating half-planes with this
* relative weight: 0.0 - only neighbors, 1.0 - same weight of the center as each neighbor.
* @param min_neibs - minimal number of connected neighbors to work with (>=2)
* @param preferDisparity - the first eigenvalue/vector is the most disparity-like
* (false - smallest eigenvalue)
* @param debugLevel debug level
* @param dbg_X supertile horizontal index to show debug information
* @param dbg_Y supertile vertical index to show debug information
* @return a pair of plane objects for each [supertile][plane][3] and a combination of both
* (to compare eigenvalues)
*/
public TilePlanes.PlaneData[][][] breakPlanesToPairs(
final TilePlanes.PlaneData[][] center_planes, // measured_planes,
final TilePlanes.PlaneData[][] neib_planes, //mod_planes,
final double center_pull,
final int min_neibs, // 2
final boolean preferDisparity, // Always start with disparity-most axis (false - lowest eigenvalue)
final int debugLevel,
final int dbg_X,
final int dbg_Y)
{
final int [][] dirsYX = {{-1, 0},{-1,1},{0,1},{1,1},{1,0},{1,-1},{0,-1},{-1,-1}};
final int tilesX = tileProcessor.getTilesX();
// final int tilesY = tileProcessor.getTilesY();
final int superTileSize = tileProcessor.getSuperTileSize();
final int stilesX = (tilesX + superTileSize -1)/superTileSize;
// final int stilesY = (tilesY + superTileSize -1)/superTileSize;
final int debug_stile = dbg_Y * stilesX + dbg_X;
final TilePlanes.PlaneData[][][] rslt_planes = new TilePlanes.PlaneData[center_planes.length][][];
final Thread[] threads = ImageDtt.newThreadArray(tileProcessor.threadsMax);
final AtomicInteger ai = new AtomicInteger(0);
final String [] titles= {"center", "first", "second", "all"};
for (int ithread = 0; ithread < threads.length; ithread++) {
threads[ithread] = new Thread() {
public void run() {
double [][] dbg_img=null;
for (int nsTile0 = ai.getAndIncrement(); nsTile0 < center_planes.length; nsTile0 = ai.getAndIncrement()) {
int sty0 = nsTile0 / stilesX;
int stx0 = nsTile0 % stilesX;
int dl = ((debugLevel > -1) && (nsTile0 == debug_stile)) ? 1:0;
if ( center_planes[nsTile0] != null) {
rslt_planes[nsTile0] = new TilePlanes.PlaneData[center_planes[nsTile0].length][];
if (dl > 0){
System.out.println("breakPlanesToPairs nsTile0="+nsTile0);
dbg_img = new double [titles.length][];
}
int np0_min = (center_planes[nsTile0].length > 1) ? 1:0; // Modify if overall plane will be removed
for (int np0 = np0_min; np0 < center_planes[nsTile0].length; np0 ++){
TilePlanes.PlaneData center_plane = center_planes[nsTile0][np0];
if (dl > 0) dbg_img[ 0] = center_plane.getPlaneDisparity(false);
int [] neibs = center_plane.getNeibBest();
int num_neibs = 0;
for (int i = 0; i < neibs.length; i++) if (neibs[i] >= 0) num_neibs++;
if (num_neibs >= min_neibs) {
// create all pairs to test
int num_splits = num_neibs * (num_neibs - 1) / 2;
double [] split_quality = new double [num_splits]; // weighted sum of eigenvalues of merged
int [][][] neibs12 = new int [2][num_splits][];
TilePlanes.PlaneData[][] plane_triads = new TilePlanes.PlaneData[num_splits][3];
int [] neib_index = new int [num_neibs];
int nn = 0;
for (int i = 0; i < neibs.length; i++) if (neibs[i] >= 0) neib_index[nn++] = i;
nn = 0;
for (int nStart = 1; nStart < num_neibs; nStart++) {
for (int nEnd = nStart+1; nEnd <= num_neibs; nEnd++){
neibs12[0][nn] = neibs.clone();
neibs12[1][nn] = neibs.clone();
for (int i = 0; i < neib_index.length; i++) {
if ((i < nStart) || (i >= nEnd)){
neibs12[0][nn][neib_index[i]] = -1;
} else {
neibs12[1][nn][neib_index[i]] = -1;
}
}
nn++;
}
}
for (int nSplit = 0; nSplit < num_splits; nSplit++) {
// Calculate merged plane for each selection (and other), weighted average and store to split_quality
for (int nis = 0; nis < 2; nis++) {
plane_triads[nSplit][nis] = center_planes[nsTile0][np0].clone();
plane_triads[nSplit][nis].setWeight(center_pull * plane_triads[nSplit][nis].getWeight());
for (int dir = 0; dir <8; dir++){
int np = neibs12[nis][nSplit][dir];
if (np >= 0){
int stx = stx0 + dirsYX[dir][1];
int sty = sty0 + dirsYX[dir][0];
int nsTile = sty * stilesX + stx; // from where to get
TilePlanes.PlaneData other_plane = plane_triads[nSplit][nis].getPlaneToThis(
neib_planes[nsTile][np],
dl); // debugLevel);
// if (dl > 0) dbg_img[ 2 + dir] = other_plane.getPlaneDisparity(false);
if (other_plane != null){
if (plane_triads[nSplit][nis].getWeight() > 0.0){
plane_triads[nSplit][nis] = plane_triads[nSplit][nis].mergePlaneToThis(
other_plane, // PlaneData otherPd,
1.0, // double scale_other,
false, // boolean ignore_weights,
preferDisparity,
dl); // int debugLevel)
} else {
plane_triads[nSplit][nis] = other_plane;
}
//if (dl > 0) dbg_img[10 + dir] = this_new_plane.getPlaneDisparity(false);
} else {
plane_triads[nSplit][nis] = null;
break;
}
}
}
} // for (int nis = 0; nis < 2; nis++) {
if ((plane_triads[nSplit][0] != null) && (plane_triads[nSplit][1] != null)){
double w1 = plane_triads[nSplit][0].getWeight();
double w2 = plane_triads[nSplit][1].getWeight();
split_quality[nSplit] = (w1 * plane_triads[nSplit][0].getValue() + w2 * plane_triads[nSplit][1].getValue())/
(w1 + w2);
} else {
split_quality[nSplit] = Double.NaN;
}
}
// find minimum in split_quality and generate a pair of plane objects, setting neighbors for each
// later use these plane pairs to assign tiles to each and generate a new eigenvalues/vectors
// TODO: How to handle a pair? Any special treatment (like fusing?),
// if the plane intersection line is inside supertile - use min/max for snapping
int best_index = -1;
for (int nSplit = 0; nSplit < num_splits; nSplit++) {
if (!Double.isNaN(split_quality[nSplit]) && ((best_index < 0) || (split_quality[nSplit] < split_quality[best_index]))){
best_index = nSplit;
}
}
if (best_index >= 0) {
plane_triads[best_index][2] = plane_triads[best_index][0].clone().mergePlaneToThis(
plane_triads[best_index][1], // PlaneData otherPd,
1.0, // double scale_other,
false, // boolean ignore_weights,
preferDisparity,
dl);
for (int nis = 0; nis < 2; nis++) {
plane_triads[best_index][nis].setNeibBest(neibs12[nis][best_index]);
}
rslt_planes[nsTile0][np0] = plane_triads[best_index];
if (dl > 0) dbg_img[ 1] = plane_triads[best_index][0].getPlaneDisparity(false);
if (dl > 0) dbg_img[ 2] = plane_triads[best_index][1].getPlaneDisparity(false);
if (dl > 0) dbg_img[ 3] = plane_triads[best_index][2].getPlaneDisparity(false);
// merged_eig_val are not set - will have to be recalculated when updating after changing tile selections
// to make each plane
// TODO save gain from splitting to planes
}
if ((dl > 0) && (debugLevel > -1)){
showDoubleFloatArrays sdfa_instance = new showDoubleFloatArrays();
sdfa_instance.showArrays(dbg_img, superTileSize, superTileSize, true, "breakPlanesToPairs"+stx0+"_y"+sty0+"_p"+np0, titles);
}
}
}
}
}
}
};
}
ImageDtt.startAndJoin(threads);
return rslt_planes;
}
public boolean [][] selectPlanes(
final double dispNorm,
final double maxEigen, // maximal eigenvalue of planes to consider
......
......@@ -204,6 +204,11 @@ public class TilePlanes {
return this.neib_best[dir];
}
public void setNeibBest(int [] vals)
{
this.neib_best = vals;
}
public void setNeibBest(int dir, int val)
{
this.neib_best[dir] = val;
......
......@@ -3038,6 +3038,17 @@ public class TileProcessor {
10.0); // double arrow_white)
// save surfaces with SuperTiles instance. They can be used to snap to for the per-tile disparity maps.
st.setSurfaces(surfaces);
TilePlanes.PlaneData[][][] split_planes =
st.breakPlanesToPairs(
st.getPlanes(), // Mod(), // final TilePlanes.PlaneData[][] center_planes, // measured_planes,
st.getPlanes(), // Mod(), // final TilePlanes.PlaneData[][] neib_planes, //mod_planes,
clt_parameters.plSplitPull , // final double center_pull,
clt_parameters.plSplitMinNeib , // min_neibs, // 2
clt_parameters.plPreferDisparity,
1, // final int debugLevel)
clt_parameters.tileX,
clt_parameters.tileY);
if (clt_parameters.show_planes){
int [] wh = st.getShowPlanesWidthHeight();
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment