Merge branch 'chroma-key-tuning-review'

.gitignore

@@ -0,0 +1,12 @@
+*.o
+*~
+*.orig
+*.a
+doc/html/
+tests/test-*
+tests/test.log
+*.rej
+ext/uncertain/
+*.obj
+build/msvc/Debug/
+build/msvc/x64/Release/

.travis.yml

@@ -19,6 +19,5 @@ addons:
       - valgrind
 
 script:
-  - make -f build/linux/Makefile.linux64
-  - ( cd tests ; make )
+  - make -f build/linux/Makefile.linux64 test
 

build/general/Makefile.inc

@@ -144,6 +144,9 @@ $(LIBRARY):	$(OBJECTS)
 		$(AR) cr $@ $^
 		$(RANLIB) $@
 
+test:		$(LIBRARY)
+		$(MAKE) -C tests
+
 depend:
 		makedepend -fbuild/general/Makefile.inc -Y -- $(CFLAGS) -- $(SOURCES)
 

dsp/chromagram/Chromagram.cpp

@@ -33,8 +33,9 @@ int Chromagram::initialise( ChromaConfig Config )
     m_BPO  = Config.BPO;		// bins per octave
     m_normalise = Config.normalise;     // if frame normalisation is required
 
-    // No. of constant Q bins
-    m_uK = (int) ceil( m_BPO * log(m_FMax/m_FMin)/log(2.0));	
+    // Extend range to a full octave
+    double octaves = log(m_FMax / m_FMin) / log(2.0);
+    m_FMax = m_FMin * pow(2.0, ceil(octaves));
 
     // Create array for chroma result
     m_chromadata = new double[ m_BPO ];
@@ -44,7 +45,7 @@ int Chromagram::initialise( ChromaConfig Config )
 
     // Populate CQ config structure with parameters
     // inherited from the Chroma config
-    ConstantQConfig.FS	 = Config.FS;
+    ConstantQConfig.FS = Config.FS;
     ConstantQConfig.min = m_FMin;
     ConstantQConfig.max = m_FMax;
     ConstantQConfig.BPO = m_BPO;
@@ -53,6 +54,9 @@ int Chromagram::initialise( ChromaConfig Config )
     // Initialise ConstantQ operator
     m_ConstantQ = new ConstantQ( ConstantQConfig );
 
+    // No. of constant Q bins
+    m_uK = m_ConstantQ->getK();
+
     // Initialise working arrays
     m_frameSize = m_ConstantQ->getfftlength();
     m_hopSize = m_ConstantQ->gethop();
@@ -121,7 +125,7 @@ void Chromagram::unityNormalise(double *src)
 }
 
 
-double* Chromagram::process( const double *data )
+double *Chromagram::process(const double *data)
 {
     if (!m_skGenerated) {
         // Generate CQ Kernel 
@@ -139,12 +143,20 @@ double* Chromagram::process( const double *data )
     }
     m_window->cut(m_windowbuf);
 
+    // The frequency-domain version expects pre-fftshifted input - so
+    // we must do the same here
+    for (int i = 0; i < m_frameSize/2; ++i) {
+        double tmp = m_windowbuf[i];
+        m_windowbuf[i] = m_windowbuf[i + m_frameSize/2];
+        m_windowbuf[i + m_frameSize/2] = tmp;
+    }
+
     m_FFT->forward(m_windowbuf, m_FFTRe, m_FFTIm);
 
     return process(m_FFTRe, m_FFTIm);
 }
 
-double* Chromagram::process( const double *real, const double *imag )
+double *Chromagram::process(const double *real, const double *imag)
 {
     if (!m_skGenerated) {
         // Generate CQ Kernel 
@@ -159,8 +171,8 @@ double* Chromagram::process( const double *real, const double *imag )
     m_ConstantQ->process( real, imag, m_CQRe, m_CQIm );
 
     // add each octave of cq data into Chromagram
-    const int octaves = (int)floor(double( m_uK/m_BPO))-1;
-    for (int octave = 0; octave <= octaves; octave++) 
+    const int octaves = m_uK / m_BPO;
+    for (int octave = 0; octave < octaves; octave++) 
     {
 	int firstBin = octave*m_BPO;
 	for (int i = 0; i < m_BPO; i++) 

dsp/chromagram/Chromagram.h

@@ -20,8 +20,8 @@
 #include "base/Window.h"
 #include "ConstantQ.h"
 
-struct ChromaConfig{
-    int FS;
+struct ChromaConfig {
+    double FS;
     double min;
     double max;
     int BPO;
@@ -35,10 +35,35 @@ class Chromagram
 public:	
     Chromagram( ChromaConfig Config );
     ~Chromagram();
-
-    double* process( const double *data ); // time domain
-    double* process( const double *real, const double *imag ); // frequency domain
-    void unityNormalise( double* src );
+
+    /**
+     * Process a time-domain input signal of length equal to
+     * getFrameSize().
+     * 
+     * The returned buffer contains the chromagram values indexed by
+     * bin, with the number of values corresponding to the BPO field
+     * in the ChromaConfig supplied at construction. It is owned by
+     * the Chromagram object and is reused from one process call to
+     * the next.
+     */
+    double *process(const double *data);
+
+    /**
+     * Process a frequency-domain input signal generated from a
+     * time-domain signal of length equal to getFrameSize() that has
+     * been windowed and "fftshifted" to place the zero index in the
+     * centre of the frame. The real and imag buffers must each
+     * contain the full getFrameSize() frequency bins.
+     * 
+     * The returned buffer contains the chromagram values indexed by
+     * bin, with the number of values corresponding to the BPO field
+     * in the ChromaConfig supplied at construction. It is owned by
+     * the Chromagram object and is reused from one process call to
+     * the next.
+     */
+    double *process(const double *real, const double *imag);
+
+    void unityNormalise(double* src);
 
     // Complex arithmetic
     double kabs( double real, double imag );

dsp/chromagram/ConstantQ.cpp

@@ -128,6 +128,9 @@ void ConstantQ::sparsekernel()
 	// Computing a hamming window
 	const unsigned hammingLength = (int) ceil( m_dQ * m_FS / ( m_FMin * pow(2,((double)(k))/(double)m_BPO)));
 
+//        cerr << "k = " << k << ", q = " << m_dQ << ", m_FMin = " << m_FMin << ", hammingLength = " << hammingLength << " (rounded up from " << (m_dQ * m_FS / ( m_FMin * pow(2,((double)(k))/(double)m_BPO))) << ")" << endl;
+
+
         unsigned origin = m_FFTLength/2 - hammingLength/2;
 
 	for (unsigned i=0; i<hammingLength; i++) 
@@ -159,7 +162,7 @@ void ConstantQ::sparsekernel()
 	    const double squaredBin = squaredModule( transfHammingWindowRe[ j ], transfHammingWindowIm[ j ]);
 	    if (squaredBin <= squareThreshold) continue;
 
-	    // Insert non-zero position indexes, doubled because they are floats
+	    // Insert non-zero position indexes
 	    sk->is.push_back(j);
 	    sk->js.push_back(k);
 
@@ -271,6 +274,7 @@ double* ConstantQ::process( const double* fftdata )
     {
 	const unsigned row = cqbin[i];
 	const unsigned col = fftbin[i];
+        if (col == 0) continue;
 	const double & r1  = real[i];
 	const double & i1  = imag[i];
 	const double & r2  = fftdata[ (2*m_FFTLength) - 2*col - 2 ];
@@ -300,7 +304,7 @@ void ConstantQ::initialise( CQConfig Config )
     // work out length of fft required for this constant Q Filter bank
     m_FFTLength = (int) pow(2, nextpow2(ceil( m_dQ*m_FS/m_FMin )));
 
-    m_hop = m_FFTLength/8; // <------ hop size is window length divided by 32
+    m_hop = m_FFTLength/8;
 
 //    std::cerr << "ConstantQ::initialise: -> fft length = " << m_FFTLength << ", hop = " << m_hop << std::endl;
 
@@ -340,10 +344,11 @@ void ConstantQ::process(const double *FFTRe, const double* FFTIm,
     {
 	const unsigned row = cqbin[i];
 	const unsigned col = fftbin[i];
+        if (col == 0) continue;
 	const double & r1  = real[i];
 	const double & i1  = imag[i];
-	const double & r2  = FFTRe[ m_FFTLength - col - 1 ];
-	const double & i2  = FFTIm[ m_FFTLength - col - 1 ];
+	const double & r2  = FFTRe[ m_FFTLength - col ];
+	const double & i2  = FFTIm[ m_FFTLength - col ];
 	// add the multiplication
 	CQRe[ row ] += (r1*r2 - i1*i2);
 	CQIm[ row ] += (r1*i2 + i1*r2);

dsp/chromagram/ConstantQ.h

@@ -20,8 +20,8 @@
 #include "maths/MathAliases.h"
 #include "maths/MathUtilities.h"
 
-struct CQConfig{
-    unsigned int FS;   // samplerate
+struct CQConfig {
+    double FS;         // samplerate
     double min;        // minimum frequency
     double max;        // maximum frequency
     unsigned int BPO;  // bins per octave
@@ -58,7 +58,7 @@ class ConstantQ {
     void deInitialise();
 
     double* m_CQdata;
-    unsigned int m_FS;
+    double m_FS;
     double m_FMin;
     double m_FMax;
     double m_dQ;