void NoveltyCurve::compute() {
const vector<vector<Real> >& frequencyBands = _frequencyBands.get();
vector<Real>& novelty = _novelty.get();
if (frequencyBands.empty())
throw EssentiaException("NoveltyCurve::compute, cannot compute from an empty input matrix");
int nFrames = frequencyBands.size();
int nBands = (int)frequencyBands[0].size();
//vector<Real> weights = weightCurve(nBands);
novelty.resize(nFrames-1);
fill(novelty.begin(), novelty.end(), Real(0.0));
vector<vector<Real> > t_frequencyBands = essentia::transpose(frequencyBands); // [bands x frames]
vector<vector<Real> > noveltyBands(nBands);
int meanSize = int(0.1 * _frameRate); // integral number of frames in 2*0.05 second
// compute novelty for each sub-band
meanSize += (meanSize % 2); // force even size // TODO: why?
for (int bandIdx=0; bandIdx<nBands; bandIdx++) {
noveltyBands[bandIdx] = noveltyFunction(t_frequencyBands[bandIdx], 1000, meanSize);
}
/////////////////////////////////////////////////////////////////////////////
// TODO: By trial-&-error I found that combining weightings (flat, quadratic,
// linear and inverse quadratic) was giving better results. Should this be
// left as is or should we allow the algorithm to work with the given
// weightings from the configuration. This overrides the parameters, so if
// left as is, they should be removed as well.
/////////////////////////////////////////////////////////////////////////////
_type = FLAT;
vector<Real> aweights = weightCurve(nBands);
_type = QUADRATIC;
vector<Real> bweights = weightCurve(nBands);
_type = LINEAR;
vector<Real> cweights = weightCurve(nBands);
_type = INVERSE_QUADRATIC;
vector<Real> dweights = weightCurve(nBands);
//sum novelty on all bands (weighted) to get a single novelty value per frame
noveltyBands = essentia::transpose(noveltyBands); // back to [frames x bands]
vector<Real> bnovelty(nFrames-1, 0.0);
vector<Real> cnovelty(nFrames-1, 0.0);
vector<Real> dnovelty(nFrames-1, 0.0);
for (int frameIdx=0; frameIdx<nFrames-1; frameIdx++) { // nFrames -1 as noveltyBands is a derivative whose size is nframes-1
const vector<Real>& frame = noveltyBands[frameIdx];
for (int bandIdx=0; bandIdx<nBands; bandIdx++) {
novelty[frameIdx] += aweights[bandIdx] * frame[bandIdx];
bnovelty[frameIdx] += bweights[bandIdx] * frame[bandIdx];
cnovelty[frameIdx] += cweights[bandIdx] * frame[bandIdx];
dnovelty[frameIdx] += dweights[bandIdx] * frame[bandIdx];
}
}
for (int frameIdx=0; frameIdx<nFrames-1; frameIdx++) {
novelty[frameIdx] *= bnovelty[frameIdx];
novelty[frameIdx] *= cnovelty[frameIdx];
novelty[frameIdx] *= dnovelty[frameIdx];
}
Algorithm * mavg = AlgorithmFactory::create("MovingAverage", "size", meanSize);
vector<Real> novelty_ma;
mavg->input("signal").set(novelty);
mavg->output("signal").set(novelty_ma);
mavg->compute();
delete mavg;
novelty.assign(novelty_ma.begin(), novelty_ma.end());
}
void NoveltyCurve::compute() {
const vector<vector<Real> >& frequencyBands = _frequencyBands.get();
vector<Real>& novelty = _novelty.get();
if (frequencyBands.empty())
throw EssentiaException("NoveltyCurve::compute, cannot compute from an empty input matrix");
int nFrames = frequencyBands.size();
int nBands = (int)frequencyBands[0].size();
//vector<Real> weights = weightCurve(nBands);
novelty.resize(nFrames-1);
fill(novelty.begin(), novelty.end(), Real(0.0));
vector<vector<Real> > t_frequencyBands = essentia::transpose(frequencyBands); // [bands x frames]
vector<vector<Real> > noveltyBands(nBands);
int meanSize = int(0.1 * _frameRate); // integral number of frames in 2*0.05 second
// compute novelty for each sub-band
meanSize += (meanSize % 2); // force even size // TODO: why?
for (int bandIdx=0; bandIdx<nBands; bandIdx++) {
noveltyBands[bandIdx] = noveltyFunction(t_frequencyBands[bandIdx], 1000, meanSize);
}
//sum novelty on all bands (weighted) to get a single novelty value per frame
noveltyBands = essentia::transpose(noveltyBands); // back to [frames x bands]
// TODO: weight curves should be pre-computed in configure() method
if (_type == HYBRID) {
// EAylon: By trial-&-error I found that combining weightings (flat, quadratic,
// linear and inverse quadratic) was giving better results.
vector<Real> aweights = weightCurve(nBands, FLAT);
vector<Real> bweights = weightCurve(nBands, QUADRATIC);
vector<Real> cweights = weightCurve(nBands, LINEAR);
vector<Real> dweights = weightCurve(nBands, INVERSE_QUADRATIC);
vector<Real> bnovelty(nFrames-1, 0.0);
vector<Real> cnovelty(nFrames-1, 0.0);
vector<Real> dnovelty(nFrames-1, 0.0);
for (int frameIdx=0; frameIdx<nFrames-1; frameIdx++) { // noveltyBands is a derivative whose size is nframes-1
for (int bandIdx=0; bandIdx<nBands; bandIdx++) {
novelty[frameIdx] += aweights[bandIdx] * noveltyBands[frameIdx][bandIdx];
bnovelty[frameIdx] += bweights[bandIdx] * noveltyBands[frameIdx][bandIdx];
cnovelty[frameIdx] += cweights[bandIdx] * noveltyBands[frameIdx][bandIdx];
dnovelty[frameIdx] += dweights[bandIdx] * noveltyBands[frameIdx][bandIdx];
}
}
for (int frameIdx=0; frameIdx<nFrames-1; frameIdx++) {
// TODO why multiplication instead of sum (or mean)?
novelty[frameIdx] *= bnovelty[frameIdx];
novelty[frameIdx] *= cnovelty[frameIdx];
novelty[frameIdx] *= dnovelty[frameIdx];
}
}
else {
// TODO weight curve should be pre-computed in configure() method
vector<Real> weights = weightCurve(nBands, _type);
for (int frameIdx=0; frameIdx<nFrames-1; frameIdx++) {
for (int bandIdx=0; bandIdx<nBands; bandIdx++) {
novelty[frameIdx] += weights[bandIdx] * noveltyBands[frameIdx][bandIdx];
}
}
}
// smoothing
Algorithm * mavg = AlgorithmFactory::create("MovingAverage", "size", meanSize);
vector<Real> novelty_ma;
mavg->input("signal").set(novelty);
mavg->output("signal").set(novelty_ma);
mavg->compute();
delete mavg;
novelty.assign(novelty_ma.begin(), novelty_ma.end());
}
