// continues to rotate 'pts1' by 'step' degrees as long as points become ever-closer
// in path-distance to pts2. the initial distance is given by D. the best distance
// is returned in array[0], while the angle at which it was achieved is in array[1].
// array[3] contains the number of calls to PathDistance.
private double[] HillClimbSearch(List <PointF> pts1, List <PointF> pts2, double D, double step)
{
double i = 0.0;
double theta = 0.0;
double d = D;
do
{
D = d; // the last angle tried was better still
theta += step;
List <PointF> newPoints = GeotrigEx.RotatePoints(pts1, GeotrigEx.Degrees2Radians(theta));
d = PathDistance(newPoints, pts2);
i++;
}while (d <= D);
return(new double[3] {
D, theta - step, i
}); // distance, angle, calls to pathdist
}
/// <summary>
///
/// </summary>
/// <param name="timepoints"></param>
/// <param name="protractor">false</param>
/// <returns></returns>
public NBestList Recognize(List <TimePointF> timepoints, bool protractor) // candidate points
{
double I = GeotrigEx.PathLength(timepoints) / (NumPoints - 1); // interval distance between points
List <PointF> points = TimePointF.ConvertList(SeriesEx.ResampleInSpace(timepoints, I));
double radians = GeotrigEx.Angle(GeotrigEx.Centroid(points), points[0], false);
points = GeotrigEx.RotatePoints(points, -radians);
points = GeotrigEx.ScaleTo(points, SquareSize);
points = GeotrigEx.TranslateTo(points, Origin, true);
List <double> vector = Unistroke.Vectorize(points); // candidate's vector representation
NBestList nbest = new NBestList();
foreach (Unistroke u in _gestures.Values)
{
if (protractor) // Protractor extension by Yang Li (CHI 2010)
{
double[] best = OptimalCosineDistance(u.Vector, vector);
double score = 1.0 / best[0];
nbest.AddResult(u.Name, score, best[0], best[1]); // name, score, distance, angle
}
else // original $1 angular invariance search -- Golden Section Search (GSS)
{
double[] best = GoldenSectionSearch(
points, // to rotate
u.Points, // to match
GeotrigEx.Degrees2Radians(-45.0), // lbound
GeotrigEx.Degrees2Radians(+45.0), // ubound
GeotrigEx.Degrees2Radians(2.0) // threshold
);
double score = 1.0 - best[0] / HalfDiagonal;
nbest.AddResult(u.Name, score, best[0], best[1]); // name, score, distance, angle
}
}
nbest.SortDescending(); // sort descending by score so that nbest[0] is best result
return(nbest);
}
/// <summary>
///
/// </summary>
/// <param name="pts1"></param>
/// <param name="pts2"></param>
/// <param name="writer"></param>
/// <returns></returns>
private double[] FullSearch(List <PointF> pts1, List <PointF> pts2, StreamWriter writer)
{
double bestA = 0d;
double bestD = PathDistance(pts1, pts2);
for (int i = -180; i <= +180; i++)
{
List <PointF> newPoints = GeotrigEx.RotatePoints(pts1, GeotrigEx.Degrees2Radians(i));
double d = PathDistance(newPoints, pts2);
if (writer != null)
{
writer.WriteLine("{0}\t{1:F3}", i, Math.Round(d, 3));
}
if (d < bestD)
{
bestD = d;
bestA = i;
}
}
writer.WriteLine("\nFull Search (360 rotations)\n{0:F2}{1}\t{2:F3} px", Math.Round(bestA, 2), (char)176, Math.Round(bestD, 3)); // calls, angle, distance
return(new double[3] {
bestD, bestA, 360.0
}); // distance, angle, calls to pathdist
}
public bool CreateRotationGraph(string file1, string file2, string dir, bool similar)
{
bool success = true;
StreamWriter writer = null;
XmlTextReader reader = null;
try
{
// read gesture file #1
reader = new XmlTextReader(file1);
reader.WhitespaceHandling = WhitespaceHandling.None;
reader.MoveToContent();
Unistroke g1 = ReadGesture(reader);
reader.Close();
// read gesture file #2
reader = new XmlTextReader(file2);
reader.WhitespaceHandling = WhitespaceHandling.None;
reader.MoveToContent();
Unistroke g2 = ReadGesture(reader);
// create output file for results
string outfile = String.Format("{0}\\{1}({2}, {3})_{4}.txt", dir, similar ? "o" : "x", g1.Name, g2.Name, Environment.TickCount);
writer = new StreamWriter(outfile, false, Encoding.UTF8);
writer.WriteLine("Rotated: {0} --> {1}. {2}, {3}\n", g1.Name, g2.Name, DateTime.Now.ToLongDateString(), DateTime.Now.ToLongTimeString());
// do the full 360 degree rotations
double[] full = FullSearch(g1.Points, g2.Points, writer);
// use bidirectional hill climbing to do it again
double init = PathDistance(g1.Points, g2.Points); // initial distance
double[] pos = HillClimbSearch(g1.Points, g2.Points, init, 1d);
double[] neg = HillClimbSearch(g1.Points, g2.Points, init, -1d);
double[] best = new double[3];
best = (neg[0] < pos[0]) ? neg : pos; // min distance
writer.WriteLine("\nHill Climb Search ({0} rotations)\n{1:F2}{2}\t{3:F3} px", pos[2] + neg[2] + 1, Math.Round(best[1], 2), (char)176, Math.Round(best[0], 3)); // calls, angle, distance
// use golden section search to do it yet again
double[] gold = GoldenSectionSearch(
g1.Points, // to rotate
g2.Points, // to match
GeotrigEx.Degrees2Radians(-45.0), // lbound
GeotrigEx.Degrees2Radians(+45.0), // ubound
GeotrigEx.Degrees2Radians(2.0)); // threshold
writer.WriteLine("\nGolden Section Search ({0} rotations)\n{1:F2}{2}\t{3:F3} px", gold[2], Math.Round(gold[1], 2), (char)176, Math.Round(gold[0], 3)); // calls, angle, distance
// use Protractor to do it yet again
// TODO
// for pasting into Excel
writer.WriteLine("\n{0} {1} {2:F2} {3:F2} {4:F3} {5:F3} {6} {7:F2} {8:F2} {9:F3} {10} {11:F2} {12:F2} {13:F3} {14}",
g1.Name, // rotated
g2.Name, // into
Math.Abs(Math.Round(full[1], 2)), // |angle|
Math.Round(full[1], 2), // Full Search angle
Math.Round(full[0], 3), // Full Search distance
Math.Round(init, 3), // Initial distance w/o any search
full[2], // Full Search iterations
Math.Abs(Math.Round(best[1], 2)), // |angle|
Math.Round(best[1], 2), // Bidirectional Hill Climb Search angle
Math.Round(best[0], 3), // Bidirectional Hill Climb Search distance
pos[2] + neg[2] + 1, // Bidirectional Hill Climb Search iterations
Math.Abs(Math.Round(gold[1], 2)), // |angle|
Math.Round(gold[1], 2), // Golden Section Search angle
Math.Round(gold[0], 3), // Golden Section Search distance
gold[2]); // Golden Section Search iterations
}
catch (XmlException xml)
{
Console.Write(xml.Message);
success = false;
}
catch (Exception ex)
{
Console.Write(ex.Message);
success = false;
}
finally
{
if (reader != null)
{
reader.Close();
}
if (writer != null)
{
writer.Close();
}
}
return(success);
}
/// <summary>
/// Tests an entire batch of files. See comments atop MainForm.TestBatch_Click().
/// </summary>
/// <param name="subject">Subject identification.</param>
/// <param name="speed">"fast", "medium", or "slow"</param>
/// <param name="categories">A list of gesture categories that each contain lists of prototypes (examples) within that gesture category.</param>
/// <param name="dir">The directory into which to write the output files.</param>
/// <param name="protractor">If true, uses Protractor instead of Golden Section Search.</param>
/// <returns>The two filenames of the output file if successful; null otherwise. The main results are in string[0],
/// while the detailed recognition results are in string[1].</returns>
public string[] TestBatch(string subject, string speed, List <Category> categories, string dir, bool protractor)
{
StreamWriter mw = null; // main results writer
StreamWriter dw = null; // detailed results writer
string[] filenames = new string[2];
try
{
// set up a main results file and detailed results file
int start = Environment.TickCount;
filenames[0] = String.Format("{0}\\$1({1})_main_{2}.txt", dir, protractor ? "protractor" : "gss", start); // main results (small file)
filenames[1] = String.Format("{0}\\$1({1})_nbest_{2}.txt", dir, protractor ? "protractor" : "gss", start); // recognition details (large file)
mw = new StreamWriter(filenames[0], false, Encoding.UTF8);
mw.WriteLine("Subject = {0}, Recognizer = $1, Search = {1}, Speed = {2}, StartTime(ms) = {3}", subject, protractor ? "protractor" : "gss", speed, start);
mw.WriteLine("Subject Recognizer Search Speed NumTraining GestureType RecognitionRate\n");
dw = new StreamWriter(filenames[1], false, Encoding.UTF8);
dw.WriteLine("Subject = {0}, Recognizer = $1, Search = {1}, Speed = {2}, StartTime(ms) = {3}", subject, protractor ? "protractor" : "gss", speed, start);
dw.WriteLine("Correct? NumTrain Tested 1stCorrect Pts Ms Angle : (NBestNames) [NBestScores]\n");
// determine the number of gesture categories and the number of examples in each one
int numCategories = categories.Count;
int numExamples = categories[0].NumExamples;
double totalTests = (numExamples - 1) * NumRandomTests;
// outermost loop: trains on N=1..9, tests on 10-N (for e.g., numExamples = 10)
for (int n = 1; n <= numExamples - 1; n++)
{
// storage for the final avg results for each category for this N
double[] results = new double[numCategories];
// run a number of tests at this particular N number of training examples
for (int r = 0; r < NumRandomTests; r++)
{
_gestures.Clear(); // clear any (old) loaded prototypes
// load (train on) N randomly selected gestures in each category
for (int i = 0; i < numCategories; i++)
{
int[] chosen = RandomEx.Array(0, numExamples - 1, n, true); // select N unique indices
for (int j = 0; j < chosen.Length; j++)
{
Unistroke p = categories[i][chosen[j]]; // get the prototype from this category at chosen[j]
_gestures.Add(p.Name, p); // load the randomly selected test gestures into the recognizer
}
}
// testing loop on all unloaded gestures in each category. creates a recognition
// rate (%) by averaging the binary outcomes (correct, incorrect) for each test.
for (int i = 0; i < numCategories; i++)
{
// pick a random unloaded gesture in this category for testing.
// instead of dumbly picking, first find out what indices aren't
// loaded, and then randomly pick from those.
int[] notLoaded = new int[numExamples - n];
for (int j = 0, k = 0; j < numExamples; j++)
{
Unistroke g = categories[i][j];
if (!_gestures.ContainsKey(g.Name))
{
notLoaded[k++] = j; // jth gesture in categories[i] is not loaded
}
}
int chosen = RandomEx.Integer(0, notLoaded.Length - 1); // index
Unistroke p = categories[i][notLoaded[chosen]]; // gesture to test
Debug.Assert(!_gestures.ContainsKey(p.Name));
// do the recognition!
List <PointF> testPts = GeotrigEx.RotatePoints( // spin gesture randomly
TimePointF.ConvertList(p.RawPoints),
GeotrigEx.Degrees2Radians(RandomEx.Integer(0, 359))
);
NBestList result = this.Recognize(TimePointF.ConvertList(testPts), protractor);
string category = Category.ParseName(result.Name);
int correct = (category == categories[i].Name) ? 1 : 0;
dw.WriteLine("{0} {1} {2} {3} {4} {5} {6:F1}{7} : ({8}) [{9}]",
correct, // Correct?
n, // NumTrain
p.Name, // Tested
FirstCorrect(p.Name, result.Names), // 1stCorrect
p.RawPoints.Count, // Pts
p.Duration, // Ms
Math.Round(result.Angle, 1), (char)176, // Angle tweaking :
result.NamesString, // (NBestNames)
result.ScoresString); // [NBestScores]
results[i] += correct;
}
// provide feedback as to how many tests have been performed thus far.
double testsSoFar = ((n - 1) * NumRandomTests) + r;
//TODO
//ProgressChangedEvent(this, new ProgressEventArgs(testsSoFar / totalTests)); // callback
}
//
// now create the final results for this N and write them to a file
//
for (int i = 0; i < numCategories; i++)
{
results[i] /= (double)NumRandomTests; // normalize by the number of tests at this N
Category c = (Category)categories[i];
// Subject Recognizer Search Speed NumTraining GestureType RecognitionRate
mw.WriteLine("{0} $1 {1} {2} {3} {4} {5:F3}",
subject,
protractor ? "protractor" : "gss",
speed,
n,
c.Name,
Math.Round(results[i], 3)
);
}
}
// time-stamp the end of the processing
int end = Environment.TickCount;
mw.WriteLine("\nEndTime(ms) = {0}, Minutes = {1:F2}", end, Math.Round((end - start) / 60000.0, 2));
dw.WriteLine("\nEndTime(ms) = {0}, Minutes = {1:F2}", end, Math.Round((end - start) / 60000.0, 2));
}
catch (Exception ex)
{
Console.WriteLine(ex.Message);
filenames = null;
}
finally
{
if (mw != null)
{
mw.Close();
}
if (dw != null)
{
dw.Close();
}
}
return(filenames);
}