From 2261a9a1f146cf528e2b5142b531d14428c12f2e Mon Sep 17 00:00:00 2001
From: Tim Redfern <tim@eclectronics.org>
Date: Tue, 10 Dec 2013 13:05:22 +0000
Subject: similarity segmentation working

---
 rotord/src/nodes_audio_analysis.cpp | 69 ++++++++++++++++++++++++++++---------
 1 file changed, 52 insertions(+), 17 deletions(-)

(limited to 'rotord/src/nodes_audio_analysis.cpp')

diff --git a/rotord/src/nodes_audio_analysis.cpp b/rotord/src/nodes_audio_analysis.cpp
index f077d41..faefb02 100644
--- a/rotord/src/nodes_audio_analysis.cpp
+++ b/rotord/src/nodes_audio_analysis.cpp
@@ -279,10 +279,17 @@ nned to retrieve total intensity by segment
 				for (int j=s.second.size()-1;j>0;j--){
 				seggrps[s.second[0]].second.push_back(make_pair(totalsmap[s.second[j]],s.second[j]));
 				//keep running average// should be by area?
-				seggrps[s.second[0]].first+=(totalsmap[s.second[j]]*(1.0/max(1,(int)seggrps[s.second[0]].second.size()-1)));
-				double div=seggrps[s.second[0]].second.size()==1?1.0:((double)seggrps[s.second[0]].second.size()-1/(double)seggrps[s.second[0]].second.size());
+				//seggrps[s.second[0]].first+=(totalsmap[s.second[j]]*(1.0/max(1,(int)seggrps[s.second[0]].second.size()-1)));
+				//double div=seggrps[s.second[0]].second.size()==1?1.0:((double)seggrps[s.second[0]].second.size()-1/(double)seggrps[s.second[0]].second.size());
 				//neat! this gives 1,1/2,2/3,3/4..
-				seggrps[s.second[0]].first*=div;
+				//seggrps[s.second[0]].first*=div;
+
+				//easier is to
+				double avg=0.0f;
+				for (auto p:seggrps[s.second[0]].second) avg+=p.first;
+				avg/=seggrps[s.second[0]].second.size();
+				seggrps[s.second[0]].first=avg;
+
 				seggrps.erase(seggrps.begin()+s.second[j]);
 				}
 			}
@@ -291,21 +298,49 @@ nned to retrieve total intensity by segment
 		//sort the contents by intensity
 		std::sort(seggrps.begin(),seggrps.end(),sortseggrps);
 		//possible mergers will be with groups with adjacent intensity
-		while (seggrps.size()>(int)parameters["levels"]->value){
-			//reduce similarity groups
-			//decide the best 2 to merge
-			vector<double> diffs;
-			for (int j=0;j<seggrps.size()-1;j++) diffs.push_back(seggrps[j+1].first-seggrps[j].first);
-			int smallest=0;
-			for (int j=1;j<diffs.size();j++) if (diffs[i]<diffs[smallest]) smallest=i;
-			for (int j=0;j<seggrps[smallest].second.size();j++) {
-				seggrps[smallest+1].second.push_back(seggrps[smallest].second[j]);
+		if (((int)parameters["levels"]->value)>0) {
+			while (seggrps.size()>(int)parameters["levels"]->value){
+				//reduce similarity groups
+				//decide the best 2 to merge
+				vector<double> diffs;
+				for (int j=0;j<seggrps.size()-1;j++) diffs.push_back(seggrps[j+1].first-seggrps[j].first);
+				int smallest=0;
+				for (int j=1;j<diffs.size();j++) if (diffs[j]<diffs[smallest]) smallest=j;
+				for (int j=0;j<seggrps[smallest].second.size();j++) {
+					seggrps[smallest+1].second.push_back(seggrps[smallest].second[j]);
+				}
+				//recalculate intensity average
+				double avg=0.0f;
+				for (auto p:seggrps[smallest+1].second) avg+=p.first;
+				avg/=seggrps[smallest+1].second.size();
+				seggrps[smallest+1].first=avg;
+
+				seggrps.erase(seggrps.begin()+smallest);
+			}
+			cerr<<"intensities merged, "<<seggrps.size()<<" levels remain"<<endl;
+			while (seggrps.size()<min((int)parameters["levels"]->value,(int)totalsmap.size())) {
+				//split groups
+				//calculate standard deviation of intensity variation
+				vector<double> devs;
+				for (int j=0;j<seggrps.size()-1;j++) {
+					double avg=0.0;
+					double dev=0.0;
+					for (int k=0;k<seggrps[j].second.size();k++){
+						avg+=k<seggrps[j].second[k].first;
+					}
+					avg/=seggrps[j].second.size();
+					for (int k=0;k<seggrps[j].second.size();k++){
+						dev+=pow(avg-k<seggrps[j].second[k].first,2.0);
+					}
+					dev/=seggrps[j].second.size();
+					devs.push_back(pow(dev,0.5));
+				}
+				//find group with largest standard deviation
+				int largest=0;
+				for (int j=1;j<devs.size();j++) if (devs[j]>devs[largest]) largest=j;
+				//sanity check: if there are any groups that can be split they will have larger SD than singleton groups
+				//TODO actually split group
 			}
-			seggrps.erase(seggrps.begin()+smallest);
-		}
-		cerr<<"intensities merged, "<<seggrps.size()<<" levels remain"<<endl;
-		while (seggrps.size()<(int)parameters["levels"]->value) {
-			//split groups
 		}
 
 		map<int,int> outputvalues;
-- 
cgit v1.2.3