Skip to content

Commit

Permalink
update
Browse files Browse the repository at this point in the history 
- update c version code, add final score threshold to remove most false positive patches
  • Loading branch information
@luoyetx
luoyetx committed Aug 2, 2016
1 parent 1dc8d63 commit eb74f84
Show file tree
Hide file tree
Showing 6 changed files with 259 additions and 60 deletions.
281 changes: 232 additions & 49 deletions c/jda.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -10,11 +10,20 @@
#define inline __inline
#endif

// jda global constance

#define JDA_T 5
#define JDA_K 290
#define JDA_LANDMARK_N 29
/*!
* \breif jda global constance
* \param JDA_T number of stages
* \param JDA_K number of carts every stage
* \param JDA_LANDMARK_N number of landmarks
* \param JDA_TREE_DEPTH depth of a cart
* \param JDA_TREE_LEAF_N leaf number of a cart
* \param JDA_CART_N number of total carts in the model
* \param JDA_LANDMARK_DIM dimension of landmarks
* \param JDA_LBF_N dimension of local binary feature
*/
#define JDA_T 3
#define JDA_K 540
#define JDA_LANDMARK_N 27
#define JDA_TREE_DEPTH 4
#define JDA_TREE_LEAF_N (1 << (JDA_TREE_DEPTH - 1))
#define JDA_TREE_NODE_N (JDA_TREE_LEAF_N - 1)
Expand Down Expand Up @@ -94,6 +103,7 @@ JDA_VECTOR(float);

/*! \breif jda bbox */
typedef struct {
/*! breif x, y, w, h */
int x, y, size;
} jdaBBox;

Expand All @@ -102,36 +112,51 @@ typedef float jdaShape[JDA_LANDMARK_DIM];

/*!\breif jda cart node */
typedef struct {
/*! breif scale */
int scale;
/*! breif landmark id */
int landmark1_x;
int landmark2_x;
/*! breif landmark offset to generate feature value */
float landmark1_offset_x;
float landmark1_offset_y;
float landmark2_offset_x;
float landmark2_offset_y;
/*! \breif feature threshold */
int th;
} jdaNode;

/*! \breif jda cart */
typedef struct {
/*! \breif nodes in this cart */
jdaNode nodes[JDA_TREE_NODE_N];
/*! \breif scores stored in the leaf nodes */
float score[JDA_TREE_LEAF_N];
/*! \breif score thrshold */
float th;
/*! \breif mean and std apply to the score */
float mean, std;
} jdaCart;

/*! \breif jda cascador */
typedef struct {
/*! \breif all carts in the model */
jdaCart carts[JDA_CART_N];
/*! \breif regression weights of every stage */
float ws[JDA_T][JDA_LBF_N][JDA_LANDMARK_DIM];
/*! \breif mean shape of the face */
float mean_shape[JDA_LANDMARK_DIM];
/*! \breif final score threshold */
float th;
} jdaCascador;

/*! \breif jda image */
typedef struct {
int w;
int h;
/*! \breif width and height */
int w, h;
/*! \breif step of a row in the image, usally equals to width */
int step;
/*! \breif gray image data */
unsigned char *data;
} jdaImage;

Expand Down Expand Up @@ -376,7 +401,7 @@ static jdaResult jdaInternalDetect(jdaCascador *cascador, jdaImage o, jdaImage h
score += cart->score[leaf_idx];
score = (score - cart->mean) / cart->std;
// not a face
if (score <= cart->th) goto next;
if (score < cart->th) goto next;
lbf[k] = k*JDA_TREE_LEAF_N + leaf_idx;
cart++;
}
Expand All @@ -389,6 +414,9 @@ static jdaResult jdaInternalDetect(jdaCascador *cascador, jdaImage o, jdaImage h
}
}
}
// final threshold
if (score < cascador->th) goto next;

jdaBBox bbox;
bbox.x = x; bbox.y = y;
bbox.size = win_size;
Expand Down Expand Up @@ -456,7 +484,7 @@ jdaResult jdaDetect(void *cascador, unsigned char *data, int width, int height)
* \breif serialize model from JDA
* \note JDA dump data type is double
*/
void *jdaCascadorCreate(const char *model) {
void *jdaCascadorCreateDouble(const char *model) {
FILE *fin = fopen(model, "rb");
if (!fin) return NULL;
jdaCascador *cascador = (jdaCascador*)malloc(sizeof(jdaCascador));
Expand All @@ -465,21 +493,21 @@ void *jdaCascadorCreate(const char *model) {
return NULL;
}

int t4;
double t8;
int i4;
double f8;
int t, k, i, j;
// meta
fread(&t4, sizeof(int), 1, fin);
fread(&t4, sizeof(int), 1, fin);
fread(&t4, sizeof(int), 1, fin);
fread(&t4, sizeof(int), 1, fin);
fread(&t4, sizeof(int), 1, fin);
fread(&t4, sizeof(int), 1, fin);
fread(&t4, sizeof(int), 1, fin);
fread(&i4, sizeof(int), 1, fin);
fread(&i4, sizeof(int), 1, fin);
fread(&i4, sizeof(int), 1, fin);
fread(&i4, sizeof(int), 1, fin);
fread(&i4, sizeof(int), 1, fin);
fread(&i4, sizeof(int), 1, fin);
fread(&i4, sizeof(int), 1, fin);
// mean shape
for (i = 0; i < JDA_LANDMARK_DIM; i++) {
fread(&t8, sizeof(double), 1, fin);
cascador->mean_shape[i] = (float)t8;
fread(&f8, sizeof(double), 1, fin);
cascador->mean_shape[i] = (float)f8;
}
// carts
for (t = 0; t < JDA_T; t++) {
Expand All @@ -488,51 +516,206 @@ void *jdaCascadorCreate(const char *model) {
// feature
for (i = 0; i < JDA_TREE_NODE_N; i++) {
jdaNode *node = &cart->nodes[i];
fread(&t4, sizeof(int), 1, fin);
node->scale = t4;
fread(&t4, sizeof(int), 1, fin);
node->landmark1_x = t4 << 1;
fread(&t4, sizeof(int), 1, fin);
node->landmark2_x = t4 << 1;
fread(&t8, sizeof(double), 1, fin);
node->landmark1_offset_x = (float)t8;
fread(&t8, sizeof(double), 1, fin);
node->landmark1_offset_y = (float)t8;
fread(&t8, sizeof(double), 1, fin);
node->landmark2_offset_x = (float)t8;
fread(&t8, sizeof(double), 1, fin);
node->landmark2_offset_y = (float)t8;
fread(&t4, sizeof(int), 1, fin);
node->th = t4;
fread(&i4, sizeof(int), 1, fin);
node->scale = i4;
fread(&i4, sizeof(int), 1, fin);
node->landmark1_x = i4 << 1;
fread(&i4, sizeof(int), 1, fin);
node->landmark2_x = i4 << 1;
fread(&f8, sizeof(double), 1, fin);
node->landmark1_offset_x = (float)f8;
fread(&f8, sizeof(double), 1, fin);
node->landmark1_offset_y = (float)f8;
fread(&f8, sizeof(double), 1, fin);
node->landmark2_offset_x = (float)f8;
fread(&f8, sizeof(double), 1, fin);
node->landmark2_offset_y = (float)f8;
fread(&i4, sizeof(int), 1, fin);
node->th = i4;
}
// scores
for (i = 0; i < JDA_TREE_LEAF_N; i++) {
fread(&t8, sizeof(double), 1, fin);
cart->score[i] = (float)t8;
fread(&f8, sizeof(double), 1, fin);
cart->score[i] = (float)f8;
}
// classificatio threshold
fread(&t8, sizeof(double), 1, fin);
cart->th = (float)t8;
//fread(&t8, sizeof(double), 1, fin);
//cart->mean = (float)t8;
//fread(&t8, sizeof(double), 1, fin);
//cart->std = (float)t8;
cart->mean = 0;
cart->std = 1;
fread(&f8, sizeof(double), 1, fin);
cart->th = (float)f8;
fread(&f8, sizeof(double), 1, fin);
cart->mean = (float)f8;
fread(&f8, sizeof(double), 1, fin);
cart->std = (float)f8;
}
// global regression weight
for (i = 0; i < JDA_LBF_N; i++) {
for (j = 0; j < JDA_LANDMARK_DIM; j++) {
fread(&t8, sizeof(double), 1, fin);
cascador->ws[t][i][j] = (float)t8;
fread(&f8, sizeof(double), 1, fin);
cascador->ws[t][i][j] = (float)f8;
}
}
}
fread(&t4, sizeof(int), 1, fin);
fread(&i4, sizeof(int), 1, fin);
fclose(fin);
// set final score threshold, this can be changed
cascador->th = 0;
return (void*)cascador;
}

void *jdaCascadorCreateFloat(const char *model) {
FILE *fin = fopen(model, "rb");
if (!fin) return NULL;
jdaCascador *cascador = (jdaCascador*)malloc(sizeof(jdaCascador));
if (!cascador) {
fclose(fin);
return NULL;
}

int i4;
float f4;
int t, k, i, j;
// meta
fread(&i4, sizeof(int), 1, fin);
fread(&i4, sizeof(int), 1, fin);
fread(&i4, sizeof(int), 1, fin);
fread(&i4, sizeof(int), 1, fin);
fread(&i4, sizeof(int), 1, fin);
fread(&i4, sizeof(int), 1, fin);
fread(&i4, sizeof(int), 1, fin);
// mean shape
for (i = 0; i < JDA_LANDMARK_DIM; i++) {
fread(&f4, sizeof(float), 1, fin);
cascador->mean_shape[i] = f4;
}
// carts
for (t = 0; t < JDA_T; t++) {
for (k = 0; k < JDA_K; k++) {
jdaCart *cart = &cascador->carts[t*JDA_K + k];
// feature
for (i = 0; i < JDA_TREE_NODE_N; i++) {
jdaNode *node = &cart->nodes[i];
fread(&i4, sizeof(int), 1, fin);
node->scale = i4;
fread(&i4, sizeof(int), 1, fin);
node->landmark1_x = i4 << 1;
fread(&i4, sizeof(int), 1, fin);
node->landmark2_x = i4 << 1;
fread(&f4, sizeof(float), 1, fin);
node->landmark1_offset_x = f4;
fread(&f4, sizeof(float), 1, fin);
node->landmark1_offset_y = f4;
fread(&f4, sizeof(float), 1, fin);
node->landmark2_offset_x = f4;
fread(&f4, sizeof(float), 1, fin);
node->landmark2_offset_y = f4;
fread(&i4, sizeof(int), 1, fin);
node->th = i4;
}
// scores
for (i = 0; i < JDA_TREE_LEAF_N; i++) {
fread(&f4, sizeof(float), 1, fin);
cart->score[i] = f4;
}
// classificatio threshold
fread(&f4, sizeof(float), 1, fin);
cart->th = f4;
fread(&f4, sizeof(float), 1, fin);
cart->mean = f4;
fread(&f4, sizeof(float), 1, fin);
cart->std = f4;
}
// global regression weight
for (i = 0; i < JDA_LBF_N; i++) {
for (j = 0; j < JDA_LANDMARK_DIM; j++) {
fread(&f4, sizeof(float), 1, fin);
cascador->ws[t][i][j] = f4;
}
}
}
fread(&i4, sizeof(int), 1, fin);
fclose(fin);
// set final score threshold, this can be changed
cascador->th = 0;
return (void*)cascador;
}

/*!
* \breif serialize model to a binary file
* \note this function serialze float data type, can reduce model size
*/
void jdaCascadorSerializeTo(void *cascador_, const char *model) {
FILE *fout = fopen(model, "wb");
if (!fout) return;
jdaCascador *cascador = (jdaCascador*)cascador_;
int i4;
float f4;
int t, k, i, j;
// meta
i4 = 0; // mask
fwrite(&i4, sizeof(int), 1, fout);
i4 = JDA_T;
fwrite(&i4, sizeof(int), 1, fout);
i4 = JDA_K;
fwrite(&i4, sizeof(int), 1, fout);
i4 = JDA_LANDMARK_N;
fwrite(&i4, sizeof(int), 1, fout);
i4 = JDA_TREE_DEPTH;
fwrite(&i4, sizeof(int), 1, fout);
i4 = JDA_T + 1;
fwrite(&i4, sizeof(int), 1, fout);
i4 = -1;
fwrite(&i4, sizeof(int), 1, fout);
// mean shape
fwrite(cascador->mean_shape, sizeof(float), JDA_LANDMARK_DIM, fout);
// carts
for (t = 0; t < JDA_T; t++) {
for (k = 0; k < JDA_K; k++) {
jdaCart *cart = &cascador->carts[t*JDA_K + k];
// feature
for (i = 0; i < JDA_TREE_NODE_N; i++) {
jdaNode *node = &cart->nodes[i];
i4 = node->scale;
fwrite(&i4, sizeof(int), 1, fout);
i4 = node->landmark1_x >> 1;
fwrite(&i4, sizeof(int), 1, fout);
i4 = node->landmark2_x >> 1;
fwrite(&i4, sizeof(int), 1, fout);
f4 = node->landmark1_offset_x;
fwrite(&f4, sizeof(float), 1, fout);
f4 = node->landmark1_offset_y;
fwrite(&f4, sizeof(float), 1, fout);
f4 = node->landmark2_offset_x;
fwrite(&f4, sizeof(float), 1, fout);
f4 = node->landmark2_offset_y;
fwrite(&f4, sizeof(float), 1, fout);
i4 = node->th;
fwrite(&i4, sizeof(int), 1, fout);
}
// scores
for (i = 0; i < JDA_TREE_LEAF_N; i++) {
f4 = cart->score[i];
fwrite(&f4, sizeof(float), 1, fout);
}
// classificatio threshold
f4 = cart->th;
fwrite(&f4, sizeof(float), 1, fout);
f4 = cart->mean;
fwrite(&f4, sizeof(float), 1, fout);
f4 = cart->std;
fwrite(&f4, sizeof(float), 1, fout);
}
// global regression weight
for (i = 0; i < JDA_LBF_N; i++) {
for (j = 0; j < JDA_LANDMARK_DIM; j++) {
f4 = cascador->ws[t][i][j];
fwrite(&f4, sizeof(float), 1, fout);
}
}
}
i4 = 0; // mask
fwrite(&i4, sizeof(int), 1, fout);
fclose(fout);
}

void jdaCascadorRelease(void *cascador) {
if (cascador) free((jdaCascador*)cascador);
}
Expand Down
Loading

0 comments on commit eb74f84

Please sign in to comment.