Loading src/main/java/com/elphel/imagej/tileprocessor/MultisceneLY.java +118 −28 Original line number Diff line number Diff line Loading @@ -22,7 +22,15 @@ public class MultisceneLY { this.threadsMax = threadsMax; this.updateStatus = updateStatus; } /** * Remeasure scene (from data in the model directory) to find number of correlation maximums * in each tile. Needed to keep only single-maximum tiles during Lazy Eye adjustment * @param clt_parameters parameters * @param lma_only keep only tiles that have LMA results * @param scene scene data * @param debug_level debug level * @return number of correlation maximums per tile (now 0,1, or 2) in linescan order */ public int [] getNumCorrMax( CLTParameters clt_parameters, boolean lma_only, Loading Loading @@ -57,6 +65,28 @@ public class MultisceneLY { return num_corr_max; } /** * Process multiscene combined depth map to extract infinity tiles * by analyzing depth map histogram * @param tp tile processor instance * @param composite_ds [tile]{disparity, strength} - per-tile disparity/strength pair * @param far_inf far limit for a tile considered to be at infinity (typical -0.5) * @param near_inf near limit for infinity tiles (typical +0.5) * @param far_fract mode - fraction of all pixels in a histogram, typical 0.05. * 0.0 corresponds to exact minimum of all tiles far_inf<=disparity<=near_inf, * 1.0 - exact maximum. 0.5 - median. * @param inf_range_offs - offset (add to) the mode value by this parameter to get average * expected disparity of objects at infinity. The result is returned in * inf_avg[0] that should be initialized to int[1] by the caller. * @param inf_range full range (symmetrical around offset mode value) of potentially * infinity tiles * @param min_inf_str minimal strength of infinity tiles * @param min_fg_str minimal strength of non-infinity tiles * @param inf_avg (should be initialized as double[1] - return parameter of average disparity * at infinity * @param debug generate debug images if true. * @return per-tile array, where true is for infinity tiles */ public static boolean [] getComboInfinity( TileProcessor tp, double [][] composite_ds, Loading Loading @@ -168,6 +198,17 @@ public class MultisceneLY { } return is_inf; } /** * Map reference scene (the last one in a sequence) disparity map and map of * the infinity tiles to all the scenes. * @param scenes sequence of scenes * @param inf_disp_ref expected disparity at infinity * @param infinity_ref map of infinity tiles for the reference (last) scene * @param debug show debug images * @param threadsMax maximal number of threads to use * @return boolean array per tile - which tiles are infinity - per scene, per tile */ public static boolean [][] infinityPerScene( QuadCLT [] scenes, double inf_disp_ref, // average disparity at infinity for ref scene Loading Loading @@ -277,6 +318,36 @@ public class MultisceneLY { } return inf_scenes; } /** * Filter tiles and generate target disparities array with NaN for the removed tiles. * Originally planned to generate both infinity tiles with target disparity equal to * the common expected infinity value and non-infinity ones with their own target * disparities, but later decided to do that in 2 passes - first processing only * infinity tiles (much smaller number), second pass - treat all as non-infinity * (own disparity) to get more data for LY offsets. * @param clust_size cluster size * @param inf_range full range for the acceptable infinity tiles centered at inf_disp_ref * @param scene_range disparity range for non-infinity in the same cluster in the same scene * @param min_num_inf minimal number of tiles (in all scenes total) in an infinity cluster * @param scenes sequence of scenes ordered by increasing timestamps * @param valid_tile [scene_index][tile_index] array indicating tiles with exactly one * correlation maximum and valid LMA * @param inf_disp_ref average disparity at infinity for the reference (last) scene * @param is_infinity which tile of which scene is infinity. May be null (all non-infinity), * if not - may be infinity from the composite depth map * @param only_infinity - only generate tiles for infinity, skip non-infinity (set to NaN) * @param in_num_tiles int array of total number of clusters or null. If not null, will return * number of used tiles in each cluster * @param in_inf_cluster boolean array of total number of clusters or null. If not null will * return if the cluster is infinity (all non-infinity tiles will be * removed from it). * @param threadsMax maximal number of threads to use * @param debug show debug images * @return target disparity per scene per tile. Uses inf_disp_ref for infinity tiles and disparity * maps for non-ifinity tiles (restored from the model directory to the scenes array). * All unused (filtered out) tiles have NaN for disparity. */ public static double [][] useTilesLY( final int clust_size, final double inf_range, // full range centered at inf_disp_ref to be used as infinity Loading Loading @@ -320,7 +391,6 @@ public class MultisceneLY { for (int ithread = 0; ithread < threads.length; ithread++) { threads[ithread] = new Thread() { public void run() { // Cluster: for (int nClust = ai.getAndIncrement(); nClust < clusters; nClust = ai.getAndIncrement()) { int clustX = nClust % clustersX; int clustY = nClust / clustersX; Loading @@ -336,8 +406,6 @@ public class MultisceneLY { valid_tile[nscene][nTile] && // next may be NaN (Math.abs(target_disparities[nscene][nTile] - inf_disp_ref) <= inf_hrange)) { num_inf_cluster[nClust]++; // inf_cluster[nClust] = true; // continue Cluster; } } } Loading @@ -364,7 +432,6 @@ public class MultisceneLY { } ImageDtt.startAndJoin(threads); // Mark suitable infinity tiles (they are NaN now) ai.set(0); for (int ithread = 0; ithread < threads.length; ithread++) { Loading Loading @@ -481,6 +548,27 @@ public class MultisceneLY { return rslt_disparities; } /** * Calculate Lazy Eye data in 2 passes - one for infinity adjustment, another - for LY. * @param clt_parameters parameters of calculation. * @param clust_size cluster size. * @param inf_range full range for the acceptable infinity tiles centered at inf_disp_ref. * @param scene_range disparity range for non-infinity in the same cluster in the same scene. * @param min_num_inf minimal number of tiles (in all scenes total) in an infinity cluster. * @param scenes sequence of scenes ordered by increasing timestamps. * @param valid_tile [scene_index][tile_index] array indicating tiles with exactly one * correlation maximum and valid LMA. * @param inf_disp_ref average disparity at infinity for the reference (last) scene * @param is_scene_infinity which tile of which scene is infinity. * @param target_disparities if not null, should be initialized to double[2][][], * will return target disparities array used for infinity [0][][] and non-infinity [1][][]. * @param dbg_disparity_offset - add to all disparities (for testing). * @param in_num_tiles null or int[2][] - will return number of tiles per cluster for * infinity [0][] and non-infinity - [1][]. * @param threadsMax maximal number of threads to use. * @param debug_level debug level, will show debug images if >-3. * @return [2][clusters][LY_sclices] LY data for infinity [0][][] and non-infinity[1][][]. */ public static double [][][] getLYDataInfNoinf( final CLTParameters clt_parameters, final int clust_size, Loading Loading @@ -584,6 +672,19 @@ public class MultisceneLY { return inf_noinf_lazy_eye_data; } /** * Calculate Lazy Eye data (now is called twice - separately for infinity and non-infinity) * @param clt_parameters parameters of calculation. * @param clust_size cluster size. * @param scenes sequence of scenes ordered by increasing timestamps. * @param target_disparities - per scene, per tile. NaN for unused tiles * @param num_tiles null or number of tiles per cluster, used to multiply strength in the LY results * @param threadsMax maximal number of threads * @param debug_level debug level. Generates debug images if >-3. * @return Lazy Eye data [clusters][LY_sclices] */ public static double [][] getLYData( final CLTParameters clt_parameters, final int clust_size, Loading @@ -596,7 +697,6 @@ public class MultisceneLY { QuadCLT last_scene = scenes[last_scene_index]; int numSens = last_scene.getNumSensors(); final int num_scenes = scenes.length; // final double[][] target_disparities = new double [num_scenes][]; final int tilesX = last_scene.tp.getTilesX(); final int tilesY = last_scene.tp.getTilesY(); final int clustersX = (int) Math.ceil(1.0 * tilesX / clust_size); Loading Loading @@ -812,14 +912,6 @@ public class MultisceneLY { for (int i = ind_length; i < titles.length; i++) { titles[i] = "combo-"+(i - ind_length); } /* (new ShowDoubleFloatArrays()).showArrays( // out of boundary 15 accum_2d_img, tilesX*(2*image_dtt.transform_size), tilesY*(2*image_dtt.transform_size), true, last_scene.getImageName()+"-CORR-ACCUM"+num_scenes, titles); */ (new ShowDoubleFloatArrays()).showArrays( // out of boundary 15 accum_2d_decimated, wh[0], Loading Loading @@ -921,14 +1013,19 @@ public class MultisceneLY { throw new IllegalArgumentException ("CPU version not yet supported"); } // Combine tiles of each cluster, assign to top-left tile of each cluster. Remember to offset by 0.5*cluster_size, // 0.5*cluster_size return lazy_eye_data; } /** * Calculate average pX, pY, Disparity (pXpYD) for each cluster to be used in LY data. * @param clust_size cluster size * @param tp_tasks TpTask array to read center positions and disparity * @param tile_clust_weights [tilesY][tilesX] - tile weights within each cluster * (sum of the values for each cluster should be 1.0) * @param threadsMax maximal number of threads to use * @return pXpYD array [tiles][] with only top-left corner tile of each cluster is * non-null and contain {pX, pY, Disparity} */ public static double [][] getAveragePxPyD( final int clust_size, final TpTask [][] tp_tasks, Loading @@ -952,9 +1049,6 @@ public class MultisceneLY { final AtomicInteger ai = new AtomicInteger(0); for (int iscene = 0; iscene < tp_tasks.length; iscene++) { final int nscene = iscene; // if (iscene == (tp_tasks.length-1)) { // System.out.println("getAveragePxPyD() nscene="+nscene); // } for (int ithread = 0; ithread < threads.length; ithread++) { threads[ithread] = new Thread() { public void run() { Loading @@ -977,7 +1071,7 @@ public class MultisceneLY { ImageDtt.startAndJoin(threads); ai.set(0); } //iterate by clustes, tiles in each cluster, ... //iterate by clusters, tiles in each cluster, ... for (int ithread = 00; ithread < threads.length; ithread++) { threads[ithread] = new Thread() { public void run() { Loading @@ -987,10 +1081,6 @@ public class MultisceneLY { int tileX0 = clust_size * clustX; int tileY0 = clust_size * clustY; int tile0 = tileY0 * tilesX + tileX0; // if (combo_pXpYD[tile0] == null) { // combo_pXpYD[tile0] = new double [3]; // } for (int ctY = 0; ctY < clust_size; ctY++) { int tileY = clustY * clust_size + ctY; if (tileY < tilesY) { Loading Loading
src/main/java/com/elphel/imagej/tileprocessor/MultisceneLY.java +118 −28 Original line number Diff line number Diff line Loading @@ -22,7 +22,15 @@ public class MultisceneLY { this.threadsMax = threadsMax; this.updateStatus = updateStatus; } /** * Remeasure scene (from data in the model directory) to find number of correlation maximums * in each tile. Needed to keep only single-maximum tiles during Lazy Eye adjustment * @param clt_parameters parameters * @param lma_only keep only tiles that have LMA results * @param scene scene data * @param debug_level debug level * @return number of correlation maximums per tile (now 0,1, or 2) in linescan order */ public int [] getNumCorrMax( CLTParameters clt_parameters, boolean lma_only, Loading Loading @@ -57,6 +65,28 @@ public class MultisceneLY { return num_corr_max; } /** * Process multiscene combined depth map to extract infinity tiles * by analyzing depth map histogram * @param tp tile processor instance * @param composite_ds [tile]{disparity, strength} - per-tile disparity/strength pair * @param far_inf far limit for a tile considered to be at infinity (typical -0.5) * @param near_inf near limit for infinity tiles (typical +0.5) * @param far_fract mode - fraction of all pixels in a histogram, typical 0.05. * 0.0 corresponds to exact minimum of all tiles far_inf<=disparity<=near_inf, * 1.0 - exact maximum. 0.5 - median. * @param inf_range_offs - offset (add to) the mode value by this parameter to get average * expected disparity of objects at infinity. The result is returned in * inf_avg[0] that should be initialized to int[1] by the caller. * @param inf_range full range (symmetrical around offset mode value) of potentially * infinity tiles * @param min_inf_str minimal strength of infinity tiles * @param min_fg_str minimal strength of non-infinity tiles * @param inf_avg (should be initialized as double[1] - return parameter of average disparity * at infinity * @param debug generate debug images if true. * @return per-tile array, where true is for infinity tiles */ public static boolean [] getComboInfinity( TileProcessor tp, double [][] composite_ds, Loading Loading @@ -168,6 +198,17 @@ public class MultisceneLY { } return is_inf; } /** * Map reference scene (the last one in a sequence) disparity map and map of * the infinity tiles to all the scenes. * @param scenes sequence of scenes * @param inf_disp_ref expected disparity at infinity * @param infinity_ref map of infinity tiles for the reference (last) scene * @param debug show debug images * @param threadsMax maximal number of threads to use * @return boolean array per tile - which tiles are infinity - per scene, per tile */ public static boolean [][] infinityPerScene( QuadCLT [] scenes, double inf_disp_ref, // average disparity at infinity for ref scene Loading Loading @@ -277,6 +318,36 @@ public class MultisceneLY { } return inf_scenes; } /** * Filter tiles and generate target disparities array with NaN for the removed tiles. * Originally planned to generate both infinity tiles with target disparity equal to * the common expected infinity value and non-infinity ones with their own target * disparities, but later decided to do that in 2 passes - first processing only * infinity tiles (much smaller number), second pass - treat all as non-infinity * (own disparity) to get more data for LY offsets. * @param clust_size cluster size * @param inf_range full range for the acceptable infinity tiles centered at inf_disp_ref * @param scene_range disparity range for non-infinity in the same cluster in the same scene * @param min_num_inf minimal number of tiles (in all scenes total) in an infinity cluster * @param scenes sequence of scenes ordered by increasing timestamps * @param valid_tile [scene_index][tile_index] array indicating tiles with exactly one * correlation maximum and valid LMA * @param inf_disp_ref average disparity at infinity for the reference (last) scene * @param is_infinity which tile of which scene is infinity. May be null (all non-infinity), * if not - may be infinity from the composite depth map * @param only_infinity - only generate tiles for infinity, skip non-infinity (set to NaN) * @param in_num_tiles int array of total number of clusters or null. If not null, will return * number of used tiles in each cluster * @param in_inf_cluster boolean array of total number of clusters or null. If not null will * return if the cluster is infinity (all non-infinity tiles will be * removed from it). * @param threadsMax maximal number of threads to use * @param debug show debug images * @return target disparity per scene per tile. Uses inf_disp_ref for infinity tiles and disparity * maps for non-ifinity tiles (restored from the model directory to the scenes array). * All unused (filtered out) tiles have NaN for disparity. */ public static double [][] useTilesLY( final int clust_size, final double inf_range, // full range centered at inf_disp_ref to be used as infinity Loading Loading @@ -320,7 +391,6 @@ public class MultisceneLY { for (int ithread = 0; ithread < threads.length; ithread++) { threads[ithread] = new Thread() { public void run() { // Cluster: for (int nClust = ai.getAndIncrement(); nClust < clusters; nClust = ai.getAndIncrement()) { int clustX = nClust % clustersX; int clustY = nClust / clustersX; Loading @@ -336,8 +406,6 @@ public class MultisceneLY { valid_tile[nscene][nTile] && // next may be NaN (Math.abs(target_disparities[nscene][nTile] - inf_disp_ref) <= inf_hrange)) { num_inf_cluster[nClust]++; // inf_cluster[nClust] = true; // continue Cluster; } } } Loading @@ -364,7 +432,6 @@ public class MultisceneLY { } ImageDtt.startAndJoin(threads); // Mark suitable infinity tiles (they are NaN now) ai.set(0); for (int ithread = 0; ithread < threads.length; ithread++) { Loading Loading @@ -481,6 +548,27 @@ public class MultisceneLY { return rslt_disparities; } /** * Calculate Lazy Eye data in 2 passes - one for infinity adjustment, another - for LY. * @param clt_parameters parameters of calculation. * @param clust_size cluster size. * @param inf_range full range for the acceptable infinity tiles centered at inf_disp_ref. * @param scene_range disparity range for non-infinity in the same cluster in the same scene. * @param min_num_inf minimal number of tiles (in all scenes total) in an infinity cluster. * @param scenes sequence of scenes ordered by increasing timestamps. * @param valid_tile [scene_index][tile_index] array indicating tiles with exactly one * correlation maximum and valid LMA. * @param inf_disp_ref average disparity at infinity for the reference (last) scene * @param is_scene_infinity which tile of which scene is infinity. * @param target_disparities if not null, should be initialized to double[2][][], * will return target disparities array used for infinity [0][][] and non-infinity [1][][]. * @param dbg_disparity_offset - add to all disparities (for testing). * @param in_num_tiles null or int[2][] - will return number of tiles per cluster for * infinity [0][] and non-infinity - [1][]. * @param threadsMax maximal number of threads to use. * @param debug_level debug level, will show debug images if >-3. * @return [2][clusters][LY_sclices] LY data for infinity [0][][] and non-infinity[1][][]. */ public static double [][][] getLYDataInfNoinf( final CLTParameters clt_parameters, final int clust_size, Loading Loading @@ -584,6 +672,19 @@ public class MultisceneLY { return inf_noinf_lazy_eye_data; } /** * Calculate Lazy Eye data (now is called twice - separately for infinity and non-infinity) * @param clt_parameters parameters of calculation. * @param clust_size cluster size. * @param scenes sequence of scenes ordered by increasing timestamps. * @param target_disparities - per scene, per tile. NaN for unused tiles * @param num_tiles null or number of tiles per cluster, used to multiply strength in the LY results * @param threadsMax maximal number of threads * @param debug_level debug level. Generates debug images if >-3. * @return Lazy Eye data [clusters][LY_sclices] */ public static double [][] getLYData( final CLTParameters clt_parameters, final int clust_size, Loading @@ -596,7 +697,6 @@ public class MultisceneLY { QuadCLT last_scene = scenes[last_scene_index]; int numSens = last_scene.getNumSensors(); final int num_scenes = scenes.length; // final double[][] target_disparities = new double [num_scenes][]; final int tilesX = last_scene.tp.getTilesX(); final int tilesY = last_scene.tp.getTilesY(); final int clustersX = (int) Math.ceil(1.0 * tilesX / clust_size); Loading Loading @@ -812,14 +912,6 @@ public class MultisceneLY { for (int i = ind_length; i < titles.length; i++) { titles[i] = "combo-"+(i - ind_length); } /* (new ShowDoubleFloatArrays()).showArrays( // out of boundary 15 accum_2d_img, tilesX*(2*image_dtt.transform_size), tilesY*(2*image_dtt.transform_size), true, last_scene.getImageName()+"-CORR-ACCUM"+num_scenes, titles); */ (new ShowDoubleFloatArrays()).showArrays( // out of boundary 15 accum_2d_decimated, wh[0], Loading Loading @@ -921,14 +1013,19 @@ public class MultisceneLY { throw new IllegalArgumentException ("CPU version not yet supported"); } // Combine tiles of each cluster, assign to top-left tile of each cluster. Remember to offset by 0.5*cluster_size, // 0.5*cluster_size return lazy_eye_data; } /** * Calculate average pX, pY, Disparity (pXpYD) for each cluster to be used in LY data. * @param clust_size cluster size * @param tp_tasks TpTask array to read center positions and disparity * @param tile_clust_weights [tilesY][tilesX] - tile weights within each cluster * (sum of the values for each cluster should be 1.0) * @param threadsMax maximal number of threads to use * @return pXpYD array [tiles][] with only top-left corner tile of each cluster is * non-null and contain {pX, pY, Disparity} */ public static double [][] getAveragePxPyD( final int clust_size, final TpTask [][] tp_tasks, Loading @@ -952,9 +1049,6 @@ public class MultisceneLY { final AtomicInteger ai = new AtomicInteger(0); for (int iscene = 0; iscene < tp_tasks.length; iscene++) { final int nscene = iscene; // if (iscene == (tp_tasks.length-1)) { // System.out.println("getAveragePxPyD() nscene="+nscene); // } for (int ithread = 0; ithread < threads.length; ithread++) { threads[ithread] = new Thread() { public void run() { Loading @@ -977,7 +1071,7 @@ public class MultisceneLY { ImageDtt.startAndJoin(threads); ai.set(0); } //iterate by clustes, tiles in each cluster, ... //iterate by clusters, tiles in each cluster, ... for (int ithread = 00; ithread < threads.length; ithread++) { threads[ithread] = new Thread() { public void run() { Loading @@ -987,10 +1081,6 @@ public class MultisceneLY { int tileX0 = clust_size * clustX; int tileY0 = clust_size * clustY; int tile0 = tileY0 * tilesX + tileX0; // if (combo_pXpYD[tile0] == null) { // combo_pXpYD[tile0] = new double [3]; // } for (int ctY = 0; ctY < clust_size; ctY++) { int tileY = clustY * clust_size + ctY; if (tileY < tilesY) { Loading
