public int CompareTo(Gesture other)
{
return(lower_bound.CompareTo(other.lower_bound));
}
public RecognitionResult Classify(Gesture candidate)
{
int feature_cnt = candidate.features.Count();
//
// calculate lower bounds for each template based on envelop
// and sort the list based on the boundaries
//
var temp_templates = templates;
foreach (var t in temp_templates)
{
float score = 1.0f;
for (int ii = 0;
ii < feature_cnt;
ii++)
{
Vector <float> fvec = candidate.features[ii];
score *= 1.0f / fvec.DotProduct(t.features[ii]);
}
t.fscore = score;
t.lower_bound = LowerBound(candidate, t);
//Debug.Log(t.gname + ":" + t.lower_bound.ToString() + ", " + t.rejection_threshold.ToString() );
}
//Debug.Log("\n");
temp_templates = temp_templates.OrderBy(x => x.lower_bound).ToList();
//
// now fully evaluate each template
//
RecognitionResult ret = new RecognitionResult();
foreach (Gesture template in temp_templates)
{
ret.tests += 1.0f;
float fscore = template.fscore * template.lower_bound;
//if (template.lower_bound > template.rejection_threshold)
if (fscore > template.rejection_threshold)
{
ret.pruned_by_rejection += 1.0f;
//continue;
}
if (fscore > ret.score)
{
ret.pruned_by_best_score += 1.0f;
//continue;
}
float score = DTW_Distance(
candidate.vecs,
template.vecs,
Gesture.r);
// if (score >= template.rejection_threshold)
// continue;
score *= template.fscore;
if (score < ret.score)
{
ret.template = template;
ret.score = score;
}
}
return(ret);
}
public void Add(Gesture alpha)
{
templates.Add(alpha);
}
public float CompareDatasets()
{
float errors = 0, errors_2D = 0, errors_3D = 0;
List <Gesture> original = templates;
List <Gesture> only2D = original.FindAll(x => x.Is2D() == true);
List <Gesture> only3D = original.FindAll(x => x.Is2D() == false);
Console.WriteLine();
Console.WriteLine("Comparing 2D");
templates = only2D;
for (int ii = 0; ii < templates.Count; ii++)
{
Gesture candidate = templates[ii];
templates.RemoveAt(ii);
var result = Classify(candidate);
if (result.template == null)
{
Console.WriteLine("Nothing " + candidate.gname);
errors_2D++;
}
else if (result.template.gname != candidate.gname)
{
Console.WriteLine(result.template.gname + " " + candidate.gname);
errors_2D++;
}
templates.Insert(ii, candidate);
}
Console.WriteLine();
Console.WriteLine("Comparing 3D");
templates = only3D;
for (int ii = 0; ii < templates.Count; ii++)
{
Gesture candidate = templates[ii];
templates.RemoveAt(ii);
var result = Classify(candidate);
if (result.template == null)
{
Console.WriteLine("Nothing " + candidate.gname);
errors_3D++;
}
else if (result.template.gname != candidate.gname)
{
Console.WriteLine(result.template.gname + " " + candidate.gname);
errors_3D++;
}
templates.Insert(ii, candidate);
}
templates = original;
Console.WriteLine();
Console.WriteLine("Comparing All");
for (int ii = 0; ii < templates.Count; ii++)
{
Gesture candidate = templates[ii];
templates.RemoveAt(ii);
var result = Classify(candidate);
if (result.template == null)
{
Console.WriteLine("Nothing " + candidate.gname);
errors++;
}
else if (result.template.gname != candidate.gname)
{
Console.WriteLine(result.template.gname + " " + candidate.gname);
errors++;
}
templates.Insert(ii, candidate);
}
Console.WriteLine();
Console.WriteLine("Error Rates:");
errors = errors / templates.Count;
errors_2D = errors_2D / only2D.Count;
errors_3D = errors_3D / only3D.Count;
Console.WriteLine("2D: \t" + errors_2D);
Console.WriteLine("3D: \t" + errors_3D);
Console.WriteLine("All:\t" + errors);
return(errors / (templates.Count));
}
public void EvaluateUserDependentPairwiseSpeakers(string path, string uname)
{
List <UserDataSet> alldata = DataLoader.LoadGestureDataFrom(path);
var temp = templates;
Console.WriteLine("CrossValidate\t\t2D\t\t3D\t\tAll");
var l2de = new List <float>();
var l3de = new List <float>();
var le = new List <float>();
var cnt = alldata[0].TrainingSamples[0].RawData[0].Count() / 33;
for (int mm = 0; mm < cnt; mm++)
{
for (int nn = mm + 1; nn < cnt; nn++)
{
foreach (UserDataSet ud in alldata)
{
if (uname == "" || ud.Path.Contains(uname))
{
templates.Clear();
foreach (GestureSample gs in ud.TrainingSamples)
{
if (gs.Gesture == GestureType.mu || gs.Gesture == GestureType.p || gs.Gesture == GestureType.q || gs.Gesture == GestureType.circle)
{
continue;
}
List <Vector <float> > data = new List <Vector <float> >();
foreach (var rd in gs.RawData)
{
var first = rd.Skip(mm * 33).Take(33).ToArray();
var second = rd.Skip(nn * 33).Take(33).ToArray();
var max = Math.Max(first.Max(), second.Max());
var merged = first.Concat(second).Select(x => x / max);
data.Add(Vector <float> .Build.DenseOfArray(merged.ToArray()));
}
Add(data, gs.Gesture.ToString());
}
Normalize();
float errors = 0, errors_2D = 0, errors_3D = 0;
List <Gesture> original = templates;
List <Gesture> only2D = original.FindAll(x => x.Is2D() == true);
List <Gesture> only3D = original.FindAll(x => x.Is2D() == false);
templates = only2D;
for (int ii = 0; ii < templates.Count; ii++)
{
Gesture candidate = templates[ii];
templates.RemoveAt(ii);
var result = Classify(candidate);
if (result.template == null)
{
errors_2D++;
templates.Insert(ii, candidate);
continue;
}
else if (result.template.gname != candidate.gname)
{
errors_2D++;
}
templates.Insert(ii, candidate);
}
templates = only3D;
for (int ii = 0; ii < templates.Count; ii++)
{
Gesture candidate = templates[ii];
templates.RemoveAt(ii);
var result = Classify(candidate);
if (result.template == null)
{
errors_3D++;
templates.Insert(ii, candidate);
continue;
}
else if (result.template.gname != candidate.gname)
{
errors_3D++;
}
templates.Insert(ii, candidate);
}
templates = original;
for (int ii = 0; ii < templates.Count; ii++)
{
Gesture candidate = templates[ii];
templates.RemoveAt(ii);
var result = Classify(candidate);
if (result.template == null)
{
errors++;
templates.Insert(ii, candidate);
continue;
}
else if (result.template.gname != candidate.gname)
{
errors++;
}
templates.Insert(ii, candidate);
}
errors = errors / templates.Count;
errors_2D = errors_2D / only2D.Count;
errors_3D = errors_3D / only3D.Count;
l2de.Add(errors_2D);
l3de.Add(errors_3D);
le.Add(errors);
templates.Clear();
}
}
templates = temp;
Console.WriteLine("Average\t" + (mm + 1) + " " + (nn + 1) + " " + l2de.Average().ToString("0.00") + "\t" + l3de.Average().ToString("0.00") + "\t" + le.Average().ToString("0.00"));
}
}
}
public void EvaluateUserDependent(string path, string uname)
{
List <UserDataSet> alldata = DataLoader.LoadGestureDataFrom(path);
var temp = templates;
Console.WriteLine("CrossValidate\t\t2D\t\t3D\t\tAll");
var l2de = new List <float>();
var l3de = new List <float>();
var le = new List <float>();
var count_map = new Dictionary <Tuple <string, string>, int>();
foreach (UserDataSet ud in alldata)
{
if (uname == "" || ud.Path.Contains(uname))
{
templates.Clear();
foreach (GestureSample gs in ud.TrainingSamples)
{
if (gs.Gesture == GestureType.mu || gs.Gesture == GestureType.p || gs.Gesture == GestureType.q || gs.Gesture == GestureType.circle)
{
continue;
}
List <Vector <float> > data = new List <Vector <float> >();
foreach (var rd in gs.RawData)
{
data.Add(Vector <float> .Build.DenseOfArray(rd));
}
Add(data, gs.Gesture.ToString());
}
Normalize();
float errors = 0, errors_2D = 0, errors_3D = 0;
List <Gesture> original = templates;
List <Gesture> only2D = original.FindAll(x => x.Is2D() == true);
List <Gesture> only3D = original.FindAll(x => x.Is2D() == false);
templates = only2D;
for (int ii = 0; ii < templates.Count; ii++)
{
Gesture candidate = templates[ii];
templates.RemoveAt(ii);
var result = Classify(candidate);
if (result.template == null)
{
errors_2D++;
templates.Insert(ii, candidate);
continue;
}
else if (result.template.gname != candidate.gname)
{
errors_2D++;
}
//count_map[Tuple.Create(candidate.gname, result.template.gname)] = count_map.ContainsKey(Tuple.Create(candidate.gname, result.template.gname))
// ? count_map[Tuple.Create(candidate.gname, result.template.gname)]+1 : 1;
templates.Insert(ii, candidate);
}
templates = only3D;
for (int ii = 0; ii < templates.Count; ii++)
{
Gesture candidate = templates[ii];
templates.RemoveAt(ii);
var result = Classify(candidate);
if (result.template == null)
{
errors_3D++;
templates.Insert(ii, candidate);
continue;
}
else if (result.template.gname != candidate.gname)
{
errors_3D++;
}
//count_map[Tuple.Create(candidate.gname, result.template.gname)] = count_map.ContainsKey(Tuple.Create(candidate.gname, result.template.gname))
// ? count_map[Tuple.Create(candidate.gname, result.template.gname)]+1 : 1;
templates.Insert(ii, candidate);
}
templates = original;
for (int ii = 0; ii < templates.Count; ii++)
{
Gesture candidate = templates[ii];
templates.RemoveAt(ii);
var result = Classify(candidate);
if (result.template == null)
{
errors++;
templates.Insert(ii, candidate);
continue;
}
else if (result.template.gname != candidate.gname)
{
errors++;
}
count_map[Tuple.Create(candidate.gname, result.template.gname)] = count_map.ContainsKey(Tuple.Create(candidate.gname, result.template.gname))
? count_map[Tuple.Create(candidate.gname, result.template.gname)] + 1 : 1;
templates.Insert(ii, candidate);
}
errors = errors / templates.Count;
errors_2D = errors_2D / only2D.Count;
errors_3D = errors_3D / only3D.Count;
l2de.Add(errors_2D);
l3de.Add(errors_3D);
le.Add(errors);
Console.WriteLine(ud.Path + "\t" + errors_2D.ToString("0.00") + "\t" + errors_3D.ToString("0.00") + "\t" + errors.ToString("0.00"));
templates.Clear();
}
}
templates = temp;
Console.WriteLine("Average\t" + l2de.Average().ToString("0.00") + "\t" + l3de.Average().ToString("0.00") + "\t" + le.Average().ToString("0.00"));
foreach (var key in count_map.OrderBy(x => x.Key.Item1).ThenBy(x => x.Key.Item2))
{
if (key.Value > 0)
{
Console.WriteLine(key);
}
}
}
