/**
* The main data extraction work is done in the constructor.
*
* This is where we:
*
* 1) Resample the trajectory to a fixed number
* of points (resample_cnt).
*
* 2) Calculate the normalized direction vector
* between each resampled point or the z-score
* normalized points.
*
* 3) Extract correction factor related data.
*
* It is highly recommended that the sample trajectory points
* have been smoothed with a low pass filter before calling
* this function. A simple exponential smoothing filter will
* probably be adequate for most case.
*/
public JackknifeFeatures(JackknifeBlades blades, List <Vector> points)
{
this.Vecs = new List <Vector>();
// Number of components per point.
int m = points[0].Size;
// resample the trajectory to a fixed number of points
this.Pts = Mathematics.Resample(points, blades.ResampleCnt);
// To track the bounding box widths,
// start with one point and expand.
Vector minimum = new Vector(Pts[0].Data);
Vector maximum = new Vector(Pts[0].Data);
// The abs distance traversed starts with zeros.
this.Abs = new Vector(0, m);
// Incrementally extract information.
for (int i = 1; i < blades.ResampleCnt; i++)
{
// In-between point direction vector.
Vector vec = Pts[i] - Pts[i - 1];
// Update correction factor features.
for (int j = 0; j < m; j++)
{
this.Abs[j] += Math.Abs(vec[j]);
minimum[j] = Math.Min(minimum[j], Pts[i][j]);
maximum[j] = Math.Max(maximum[j], Pts[i][j]);
}
// Save the points or direction vectors,
// depending on the selected measure.
if (blades.InnerProduct)
{
this.Vecs.Add(vec.Normalize());
}
else if (blades.EuclideanDistance)
{
// In ED scenario, make sure not to forget first point as
// loop starts at 1.
if (i == 1)
{
this.Vecs.Add(Pts[0]);
}
this.Vecs.Add(Pts[i]);
}
else
{
throw new Exception("This should not happen!");
}
}
// Z-score normalize the vecs if required,
// typically only if using euclidean distance
if (blades.ZNormalize)
{
Mathematics.ZNormalize(this.Vecs);
}
// normalize the correction factor vectors
this.Abs.Normalize();
this.Bb = (maximum - minimum).Normalize();
}
/**
* Learn a rejection threshold for each template.
* Call after all templates have been added.
*
* See explanation in jackknife_train to understand
* the input parameters.
*/
public void Train(int gpsrN, int gpsrR, double beta)
{
int template_cnt = templates.Count;
List <Distributions> distributions = new List <Distributions>();
List <Vector> synthetic = new List <Vector>();
double worst_score = 0.0;
Random rand = new Random();
//
// Create negative samples.
//
for (int i = 0; i < 1000; i++)
{
synthetic.Clear();
// Splice two samples together
// to create one negative sample.
for (int j = 0; j < 2; j++)
{
int t = rand.Next(template_cnt);
Sample s = templates[t].Sample;
int len = s.Trajectory.Count;
int start = rand.Next(len / 2);
for (int kk = 0; kk < len / 2; kk++)
{
synthetic.Add(s.Trajectory[start + kk]);
}
}
JackknifeFeatures features = new JackknifeFeatures(Blades, synthetic);
// and score it
for (int t = 0; t < template_cnt; t++)
{
double score = DTW(
features.Vecs,
templates[t].Features.Vecs);
if (worst_score < score)
{
worst_score = score;
}
if (i > 50)
{
distributions[t].AddNegativeScore(score);
}
}
// Generate a few samples to get an estimate
// of worst possible score.
if (i != 50)
{
continue;
}
// allocate distributions
for (int t = 0; t < template_cnt; t++)
{
distributions.Add(new Distributions(worst_score, 1000));
}
}
//
// Create positive examples.
//
for (int t = 0; t < template_cnt; t++)
{
for (int i = 0; i < 1000; i++)
{
synthetic.Clear();
// Create a synthetic variation of the sample.
synthetic = Mathematics.GPSR(templates[t].Sample.Trajectory, gpsrN, 0.25, gpsrR);
JackknifeFeatures features = new JackknifeFeatures(Blades, synthetic);
// and score it
double score = DTW(features.Vecs, templates[t].Features.Vecs);
distributions[t].AddPositiveScore(score);
}
}
//
// Now extract the rejection thresholds.
//
for (int t = 0; t < template_cnt; t++)
{
double threshold = distributions[t].RejectionThreshold(beta);
JackknifeTemplate temp = templates[t];
temp.RejectionThreshold = threshold;
templates[t] = temp;
}
}
