void ccv_sample_up(ccv_dense_matrix_t* a, ccv_dense_matrix_t** b, int type, int src_x, int src_y)
{
assert(src_x >= 0 && src_y >= 0);
ccv_declare_matrix_signature(sig, a->sig != 0, ccv_sign_with_format(64, "ccv_sample_up(%d,%d)", src_x, src_y), a->sig, 0);
type = (type == 0) ? CCV_GET_DATA_TYPE(a->type) | CCV_GET_CHANNEL(a->type) : CCV_GET_DATA_TYPE(type) | CCV_GET_CHANNEL(a->type);
ccv_dense_matrix_t* db = *b = ccv_dense_matrix_renew(*b, a->rows * 2, a->cols * 2, CCV_ALL_DATA_TYPE | CCV_GET_CHANNEL(a->type), type, sig);
ccv_matrix_return_if_cached(, db);
int ch = CCV_GET_CHANNEL(a->type);
int cols0 = a->cols - 1 - src_x;
int y, x, sy = -1 + src_y, sx = src_x * ch, k;
int* tab = (int*)alloca((a->cols + src_x + 2) * ch * sizeof(int));
for (x = 0; x < a->cols + src_x + 2; x++)
for (k = 0; k < ch; k++)
tab[x * ch + k] = ((x >= a->cols) ? a->cols * 2 - 1 - x : x) * ch + k;
unsigned char* buf = (unsigned char*)alloca(3 * db->cols * ch * ccv_max(CCV_GET_DATA_TYPE_SIZE(db->type), sizeof(int)));
int bufstep = db->cols * ch * ccv_max(CCV_GET_DATA_TYPE_SIZE(db->type), sizeof(int));
unsigned char* b_ptr = db->data.u8;
/* why src_y * 2: the same argument as in ccv_sample_down */
#define for_block(_for_get_a, _for_set, _for_get, _for_set_b) \
for (y = 0; y < a->rows; y++) \
{ \
for (; sy <= y + 1 + src_y; sy++) \
{ \
unsigned char* row = buf + ((sy + src_y * 2 + 1) % 3) * bufstep; \
int _sy = (sy < 0) ? -1 - sy : (sy >= a->rows) ? a->rows * 2 - 1 - sy : sy; \
unsigned char* a_ptr = a->data.u8 + a->step * _sy; \
if (a->cols == 1) \
{ \
for (k = 0; k < ch; k++) \
{ \
_for_set(row, k, _for_get_a(a_ptr, k, 0) * (G025 + G075 + G125), 0); \
_for_set(row, k + ch, _for_get_a(a_ptr, k, 0) * (G025 + G075 + G125), 0); \
} \
continue; \
} \
if (sx == 0) \
{ \
for (k = 0; k < ch; k++) \
{ \
_for_set(row, k, _for_get_a(a_ptr, k + sx, 0) * (G025 + G075) + _for_get_a(a_ptr, k + sx + ch, 0) * G125, 0); \
_for_set(row, k + ch, _for_get_a(a_ptr, k + sx, 0) * (G125 + G025) + _for_get_a(a_ptr, k + sx + ch, 0) * G075, 0); \
} \
} \
/* some serious flaw in computing Gaussian weighting in previous version
* specially, we are doing perfect upsampling (2x) so, it concerns a grid like:
* XXYY
* XXYY
* in this case, to upsampling, the weight should be from distance 0.25 and 1.25, and 0.25 and 0.75
* previously, it was mistakingly be 0.0 1.0, 0.5 0.5 (imperfect upsampling (2x - 1)) */ \
for (x = (sx == 0) ? ch : 0; x < cols0 * ch; x += ch) \
{ \
for (k = 0; k < ch; k++) \
{ \
_for_set(row, x * 2 + k, _for_get_a(a_ptr, x + sx - ch + k, 0) * G075 + _for_get_a(a_ptr, x + sx + k, 0) * G025 + _for_get_a(a_ptr, x + sx + ch + k, 0) * G125, 0); \
_for_set(row, x * 2 + ch + k, _for_get_a(a_ptr, x + sx - ch + k, 0) * G125 + _for_get_a(a_ptr, x + sx + k, 0) * G025 + _for_get_a(a_ptr, x + sx + ch + k, 0) * G075, 0); \
} \
} \
x_block(_for_get_a, _for_set, _for_get, _for_set_b); \
} \
static void _ccv_init_cubic_coeffs(int si, int sz, float s, ccv_cubic_coeffs_t* coeff)
{
const float A = -0.75f;
coeff->si[0] = ccv_max(si - 1, 0);
coeff->si[1] = si;
coeff->si[2] = ccv_min(si + 1, sz - 1);
coeff->si[3] = ccv_min(si + 2, sz - 1);
float x = s - si;
coeff->coeffs[0] = ((A * (x + 1) - 5 * A) * (x + 1) + 8 * A) * (x + 1) - 4 * A;
coeff->coeffs[1] = ((A + 2) * x - (A + 3)) * x * x + 1;
coeff->coeffs[2] = ((A + 2) * (1 - x) - (A + 3)) * (1 - x) * (1 - x) + 1;
coeff->coeffs[3] = 1.f - coeff->coeffs[0] - coeff->coeffs[1] - coeff->coeffs[2];
}
static void _ccv_init_cubic_integer_coeffs(int si, int sz, float s, ccv_cubic_integer_coeffs_t* coeff)
{
const float A = -0.75f;
coeff->si[0] = ccv_max(si - 1, 0);
coeff->si[1] = si;
coeff->si[2] = ccv_min(si + 1, sz - 1);
coeff->si[3] = ccv_min(si + 2, sz - 1);
float x = s - si;
const int W_BITS = 1 << 6;
coeff->coeffs[0] = (int)((((A * (x + 1) - 5 * A) * (x + 1) + 8 * A) * (x + 1) - 4 * A) * W_BITS + 0.5);
coeff->coeffs[1] = (int)((((A + 2) * x - (A + 3)) * x * x + 1) * W_BITS + 0.5);
coeff->coeffs[2] = (int)((((A + 2) * (1 - x) - (A + 3)) * (1 - x) * (1 - x) + 1) * W_BITS + 0.5);
coeff->coeffs[3] = W_BITS - coeff->coeffs[0] - coeff->coeffs[1] - coeff->coeffs[2];
}
static int _ccv_nnc_gemm_back_sse2(const ccv_nnc_tensor_view_t* const g, const ccv_nnc_tensor_view_t* const a, const ccv_nnc_tensor_view_t* const w, ccv_nnc_tensor_view_t* const dw, ccv_nnc_tensor_view_t* const bias, ccv_nnc_tensor_view_t* const h, const int flags)
{
const int* dwinc = CCV_IS_TENSOR_VIEW(dw) ? dw->inc : dw->info.dim;
if (!(flags & CCV_NNC_ACCUMULATE_OUTPUT)) // reset the gradients to 0
{
memset(dw->data.u8, 0, sizeof(float) * dwinc[1] * dw->info.dim[0]);
memset(bias->data.u8, 0, sizeof(float) * bias->info.dim[0]);
}
const int a_nd = ccv_nnc_tensor_nd(a->info.dim);
const int* adim = (a_nd == 1) ? a->info.dim : a->info.dim + 1;
const int g_nd = ccv_nnc_tensor_nd(g->info.dim);
const int* gdim = (g_nd == 1) ? g->info.dim : g->info.dim + 1;
const int batch_size = a_nd == 1 ? 1 : ccv_max(1, a->info.dim[0]);
int i, j;
float* gp = g->data.f32;
float* bp = bias->data.f32;
assert(bias->info.dim[0] == gdim[0]);
const int* ginc = CCV_IS_TENSOR_VIEW(g) ? ((g_nd == 1) ? g->inc : g->inc + 1) : gdim;
for (i = 0; i < batch_size; i++)
{
for (j = 0; j < gdim[0] - 3; j += 4)
{
__m128 g4 = _mm_load_ps(gp + j);
__m128 b4 = _mm_load_ps(bp + j);
_mm_stream_ps(bp + j, _mm_add_ps(b4, g4));
}
gp += ginc[0];
}
assert(gdim[0] == dw->info.dim[0]);
assert(adim[0] == dw->info.dim[1]);
const int* ainc = CCV_IS_TENSOR_VIEW(a) ? ((a_nd == 1) ? a->inc : a->inc + 1) : adim;
for (i = 0; i < batch_size; i++)
{
const float* const gp = g->data.f32 + i * ginc[0];
const float* const ap = a->data.f32 + i * ainc[0];
parallel_for(j, gdim[0]) {
float* const dwp = dw->data.f32 + j * dwinc[1];
__m128 g4 = _mm_set1_ps(gp[j]);
int k;
for (k = 0; k < adim[0] - 3; k+= 4)
{
__m128 a4 = _mm_load_ps(ap + k);
__m128 dw4 = _mm_load_ps(dwp + k);
_mm_stream_ps(dwp + k, _mm_add_ps(dw4, _mm_mul_ps(a4, g4)));
}
} parallel_endfor
}
static int _ccv_nnc_gemm_forw_sse2(const ccv_nnc_tensor_view_t* const a, const ccv_nnc_tensor_view_t* const w, const ccv_nnc_tensor_view_t* const bias, ccv_nnc_tensor_view_t* const b)
{
const int a_nd = ccv_nnc_tensor_nd(a->info.dim);
const int* adim = (a_nd == 1) ? a->info.dim : a->info.dim + 1;
const int b_nd = ccv_nnc_tensor_nd(b->info.dim);
const int* bdim = (b_nd == 1) ? b->info.dim : b->info.dim + 1;
assert(bdim[0] == bias->info.dim[0]);
assert(bdim[0] == w->info.dim[0]);
assert(adim[0] == w->info.dim[1]);
const int* ainc = CCV_IS_TENSOR_VIEW(a) ? (a_nd == 1 ? a->inc : a->inc + 1) : adim;
const int* binc = CCV_IS_TENSOR_VIEW(b) ? (b_nd == 1 ? b->inc : b->inc + 1) : bdim;
const int* winc = CCV_IS_TENSOR_VIEW(w) ? w->inc : w->info.dim;
const int batch_size = a_nd == 1 ? 1 : ccv_max(1, a->info.dim[0]);
int i;
for (i = 0; i < batch_size; i++)
{
const float* const ap = a->data.f32 + i * ainc[0];
float* const bp = b->data.f32 + i * binc[0];
parallel_for(j, bdim[0]) {
const float* const wp = w->data.f32 + j * winc[1];
int k;
__m128 v40 = _mm_set_ss(bias->data.f32[j]);
__m128 v41 = _mm_setzero_ps();
for (k = 0; k < adim[0] - 7; k += 8)
{
__m128 ap40 = _mm_load_ps(ap + k);
__m128 ap41 = _mm_load_ps(ap + k + 4);
__m128 w40 = _mm_load_ps(wp + k);
__m128 w41 = _mm_load_ps(wp + k + 4);
v40 =_mm_add_ps(_mm_mul_ps(w40, ap40), v40);
v41 =_mm_add_ps(_mm_mul_ps(w41, ap41), v41);
}
v40 = _mm_add_ps(v40, v41);
v41 = _mm_add_ps(v40, _mm_movehl_ps(v40, v40));
v40 = _mm_add_ss(v41, _mm_shuffle_ps(v41, v41, 1));
_mm_store_ss(bp + j, v40);
} parallel_endfor
}
return CCV_NNC_EXEC_SUCCESS;
}
ccv_array_t* ccv_array_new(int rsize, int rnum, uint64_t sig)
{
ccv_array_t* array;
if (ccv_cache_opt && sig != 0)
{
uint8_t type;
array = (ccv_array_t*)ccv_cache_out(&ccv_cache, sig, &type);
if (array)
{
assert(type == 1);
array->type |= CCV_GARBAGE;
array->refcount = 1;
return array;
}
}
array = (ccv_array_t*)ccmalloc(sizeof(ccv_array_t));
array->sig = sig;
array->type = CCV_REUSABLE & ~CCV_GARBAGE;
array->rnum = 0;
array->rsize = rsize;
array->size = ccv_max(rnum, 2 /* allocate memory for at least 2 items */);
array->data = ccmalloc((size_t)array->size * (size_t)rsize);
return array;
}
static void _ccv_set_union_mser(ccv_dense_matrix_t* a, ccv_dense_matrix_t* h, ccv_dense_matrix_t* b, ccv_array_t* seq, ccv_mser_param_t params)
{
assert(params.direction == CCV_BRIGHT_TO_DARK || params.direction == CCV_DARK_TO_BRIGHT);
int v, i, j;
ccv_mser_node_t* node = (ccv_mser_node_t*)ccmalloc(sizeof(ccv_mser_node_t) * a->rows * a->cols);
ccv_mser_node_t** rnode = (ccv_mser_node_t**)ccmalloc(sizeof(ccv_mser_node_t*) * a->rows * a->cols);
if (params.range <= 0)
params.range = 255;
// put it in a block so that the memory allocated can be released in the end
int* buck = (int*)alloca(sizeof(int) * (params.range + 2));
memset(buck, 0, sizeof(int) * (params.range + 2));
ccv_mser_node_t* pnode = node;
// this for_block is the only computation that can be shared between dark to bright and bright to dark
// two MSER alternatives, and it only occupies 10% of overall time, we won't share this computation
// at all (also, we need to reinitialize node for the two passes anyway).
if (h != 0)
{
unsigned char* aptr = a->data.u8;
unsigned char* hptr = h->data.u8;
#define for_block(_for_get_a, _for_get_h) \
for (i = 0; i < a->rows; i++) \
{ \
for (j = 0; j < a->cols; j++) \
if (!_for_get_h(hptr, j, 0)) \
++buck[_for_get_a(aptr, j, 0)]; \
aptr += a->step; \
hptr += h->step; \
} \
for (i = 1; i <= params.range; i++) \
buck[i] += buck[i - 1]; \
buck[params.range + 1] = buck[params.range]; \
aptr = a->data.u8; \
hptr = h->data.u8; \
for (i = 0; i < a->rows; i++) \
{ \
for (j = 0; j < a->cols; j++) \
{ \
_ccv_mser_init_node(pnode, j, i); \
if (!_for_get_h(hptr, j, 0)) \
rnode[--buck[_for_get_a(aptr, j, 0)]] = pnode; \
else \
pnode->shortcut = 0; /* this means the pnode is not available */ \
++pnode; \
} \
aptr += a->step; \
hptr += h->step; \
}
ccv_matrix_getter_integer_only_a(a->type, ccv_matrix_getter_integer_only, h->type, for_block);
#undef for_block
} else {
unsigned char* aptr = a->data.u8;
#define for_block(_, _for_get) \
for (i = 0; i < a->rows; i++) \
{ \
for (j = 0; j < a->cols; j++) \
++buck[_for_get(aptr, j, 0)]; \
aptr += a->step; \
} \
for (i = 1; i <= params.range; i++) \
buck[i] += buck[i - 1]; \
buck[params.range + 1] = buck[params.range]; \
aptr = a->data.u8; \
for (i = 0; i < a->rows; i++) \
{ \
for (j = 0; j < a->cols; j++) \
{ \
_ccv_mser_init_node(pnode, j, i); \
rnode[--buck[_for_get(aptr, j, 0)]] = pnode; \
++pnode; \
} \
aptr += a->step; \
}
ccv_matrix_getter_integer_only(a->type, for_block);
#undef for_block
}
ccv_array_t* history_list = ccv_array_new(sizeof(ccv_mser_history_t), 64, 0);
for (v = 0; v <= params.range; v++)
{
int range_segment = buck[params.direction == CCV_DARK_TO_BRIGHT ? v : params.range - v];
int range_segment_cap = buck[params.direction == CCV_DARK_TO_BRIGHT ? v + 1 : params.range - v + 1];
for (i = range_segment; i < range_segment_cap; i++)
{
pnode = rnode[i];
// try to merge pnode with its neighbors
static int dx[] = {-1, 0, 1, -1, 1, -1, 0, 1};
static int dy[] = {-1, -1, -1, 0, 0, 1, 1, 1};
ccv_mser_node_t* node0 = _ccv_mser_find_root(pnode);
for (j = 0; j < 8; j++)
{
int x = dx[j] + pnode->point.x;
int y = dy[j] + pnode->point.y;
if (x >= 0 && x < a->cols && y >= 0 && y < a->rows)
{
ccv_mser_node_t* nnode = pnode + dx[j] + dy[j] * a->cols;
if (nnode->shortcut == 0) // this is a void node, skip
continue;
ccv_mser_node_t* node1 = _ccv_mser_find_root(nnode);
if (node0 != node1)
{
// grep the extended root information
ccv_mser_history_t* root0 = (node0->root >= 0) ? (ccv_mser_history_t*)ccv_array_get(history_list, node0->root) : 0;
ccv_mser_history_t* root1 = (node1->root >= 0) ? (ccv_mser_history_t*)ccv_array_get(history_list, node1->root) : 0;
// swap the node if root1 has higher rank, or larger in size, or root0 is non-existent
if ((root0 && root1 && (root1->value > root0->value
|| (root1->value == root0->value && root1->rank > root0->rank)
|| (root1->value == root0->value && root1->rank == root0->rank && root1->size > root0->size)))
|| (root1 && !root0))
{
ccv_mser_node_t* exnode = node0;
node0 = node1;
node1 = exnode;
ccv_mser_history_t* root = root0;
root0 = root1;
root1 = root;
}
if (!root0)
{
ccv_mser_history_t root = {
.rank = 0,
.size = 1,
.value = v,
.shortcut = history_list->rnum,
.parent = history_list->rnum,
.head = node0,
.tail = node1
};
node0->root = history_list->rnum;
ccv_array_push(history_list, &root);
root0 = (ccv_mser_history_t*)ccv_array_get(history_list, history_list->rnum - 1);
assert(node1->root == -1);
} else if (root0->value < v) {
// conceal the old root as history (er), making a new one and pointing to it
root0->shortcut = root0->parent = history_list->rnum;
ccv_mser_history_t root = *root0;
root.value = v;
node0->root = history_list->rnum;
ccv_array_push(history_list, &root);
root0 = (ccv_mser_history_t*)ccv_array_get(history_list, history_list->rnum - 1);
root1 = (node1->root >= 0) ? (ccv_mser_history_t*)ccv_array_get(history_list, node1->root) : 0; // the memory may be reallocated
root0->rank = ccv_max(root0->rank, (root1 ? root1->rank : 0)) + 1;
}
if (root1)
{
if (root1->value < root0->value) // in this case, root1 is sealed as well
root1->parent = node0->root;
// thus, if root1->parent == itself && root1->shortcut != itself
// it is voided, and not sealed
root1->shortcut = node0->root;
}
// merge the two
node1->shortcut = node0;
root0->size += root1 ? root1->size : 1;
/* insert one endless double link list to another, see illustration:
* 0->1->2->3->4->5->0
* a->b->c->d->a
* set 5.next (0.prev.next) point to a
* set 0.prev point to d
* set d.next (a.prev.next) point to 0
* set a.prev point to 5
* the result endless double link list will be:
* 0->1->2->3->4->5->a->b->c->d->0 */
node0->prev->next = node1;
ccv_mser_node_t* prev = node0->prev;
node0->prev = node1->prev;
node1->prev->next = node0; // consider self-referencing
node1->prev = prev;
root0->head = node0;
root0->tail = node0->prev;
}
}
}
/* the following code is adopted from OpenCV cvPyrDown */
void ccv_sample_down(ccv_dense_matrix_t* a, ccv_dense_matrix_t** b, int type, int src_x, int src_y)
{
assert(src_x >= 0 && src_y >= 0);
ccv_declare_derived_signature(sig, a->sig != 0, ccv_sign_with_format(64, "ccv_sample_down(%d,%d)", src_x, src_y), a->sig, CCV_EOF_SIGN);
type = (type == 0) ? CCV_GET_DATA_TYPE(a->type) | CCV_GET_CHANNEL(a->type) : CCV_GET_DATA_TYPE(type) | CCV_GET_CHANNEL(a->type);
ccv_dense_matrix_t* db = *b = ccv_dense_matrix_renew(*b, a->rows / 2, a->cols / 2, CCV_ALL_DATA_TYPE | CCV_GET_CHANNEL(a->type), type, sig);
ccv_object_return_if_cached(, db);
int ch = CCV_GET_CHANNEL(a->type);
int cols0 = db->cols - 1 - src_x;
int dy, sy = -2 + src_y, sx = src_x * ch, dx, k;
int* tab = (int*)alloca((a->cols + src_x + 2) * ch * sizeof(int));
for (dx = 0; dx < a->cols + src_x + 2; dx++)
for (k = 0; k < ch; k++)
tab[dx * ch + k] = ((dx >= a->cols) ? a->cols * 2 - 1 - dx : dx) * ch + k;
unsigned char* buf = (unsigned char*)alloca(5 * db->cols * ch * ccv_max(CCV_GET_DATA_TYPE_SIZE(db->type), sizeof(int)));
int bufstep = db->cols * ch * ccv_max(CCV_GET_DATA_TYPE_SIZE(db->type), sizeof(int));
#ifdef __clang_analyzer__
memset(buf, 0, 5 * bufstep);
#endif
unsigned char* b_ptr = db->data.u8;
/* why is src_y * 4 in computing the offset of row?
* Essentially, it means sy - src_y but in a manner that doesn't result negative number.
* notice that we added src_y before when computing sy in the first place, however,
* it is not desirable to have that offset when we try to wrap it into our 5-row buffer (
* because in later rearrangement, we have no src_y to backup the arrangement). In
* such micro scope, we managed to stripe 5 addition into one shift and addition. */
#define for_block(_for_get_a, _for_set, _for_get, _for_set_b) \
for (dy = 0; dy < db->rows; dy++) \
{ \
for(; sy <= dy * 2 + 2 + src_y; sy++) \
{ \
unsigned char* row = buf + ((sy + src_y * 4 + 2) % 5) * bufstep; \
int _sy = (sy < 0) ? -1 - sy : (sy >= a->rows) ? a->rows * 2 - 1 - sy : sy; \
unsigned char* a_ptr = a->data.u8 + a->step * _sy; \
for (k = 0; k < ch; k++) \
_for_set(row, k, _for_get_a(a_ptr, sx + k, 0) * 10 + _for_get_a(a_ptr, ch + sx + k, 0) * 5 + _for_get_a(a_ptr, 2 * ch + sx + k, 0), 0); \
for(dx = ch; dx < cols0 * ch; dx += ch) \
for (k = 0; k < ch; k++) \
_for_set(row, dx + k, _for_get_a(a_ptr, dx * 2 + sx + k, 0) * 6 + (_for_get_a(a_ptr, dx * 2 + sx + k - ch, 0) + _for_get_a(a_ptr, dx * 2 + sx + k + ch, 0)) * 4 + _for_get_a(a_ptr, dx * 2 + sx + k - ch * 2, 0) + _for_get_a(a_ptr, dx * 2 + sx + k + ch * 2, 0), 0); \
x_block(_for_get_a, _for_set, _for_get, _for_set_b); \
} \
unsigned char* rows[5]; \
for(k = 0; k < 5; k++) \
rows[k] = buf + ((dy * 2 + k) % 5) * bufstep; \
for(dx = 0; dx < db->cols * ch; dx++) \
_for_set_b(b_ptr, dx, (_for_get(rows[2], dx, 0) * 6 + (_for_get(rows[1], dx, 0) + _for_get(rows[3], dx, 0)) * 4 + _for_get(rows[0], dx, 0) + _for_get(rows[4], dx, 0)) / 256, 0); \
b_ptr += db->step; \
}
int no_8u_type = (a->type & CCV_8U) ? CCV_32S : a->type;
if (src_x > 0)
{
#define x_block(_for_get_a, _for_set, _for_get, _for_set_b) \
for (dx = cols0 * ch; dx < db->cols * ch; dx += ch) \
for (k = 0; k < ch; k++) \
_for_set(row, dx + k, _for_get_a(a_ptr, tab[dx * 2 + sx + k], 0) * 6 + (_for_get_a(a_ptr, tab[dx * 2 + sx + k - ch], 0) + _for_get_a(a_ptr, tab[dx * 2 + sx + k + ch], 0)) * 4 + _for_get_a(a_ptr, tab[dx * 2 + sx + k - ch * 2], 0) + _for_get_a(a_ptr, tab[dx * 2 + sx + k + ch * 2], 0), 0);
ccv_matrix_getter_a(a->type, ccv_matrix_setter_getter, no_8u_type, ccv_matrix_setter_b, db->type, for_block);
#undef x_block
} else {
#define x_block(_for_get_a, _for_set, _for_get, _for_set_b) \
for (k = 0; k < ch; k++) \
_for_set(row, (db->cols - 1) * ch + k, _for_get_a(a_ptr, a->cols * ch + sx - ch + k, 0) * 10 + _for_get_a(a_ptr, (a->cols - 2) * ch + sx + k, 0) * 5 + _for_get_a(a_ptr, (a->cols - 3) * ch + sx + k, 0), 0);
ccv_matrix_getter_a(a->type, ccv_matrix_setter_getter, no_8u_type, ccv_matrix_setter_b, db->type, for_block);
#undef x_block
}
#undef for_block
}
static void _ccv_resample_area_8u(ccv_dense_matrix_t* a, ccv_dense_matrix_t* b)
{
assert(a->cols > 0 && b->cols > 0);
ccv_int_alpha* xofs = (ccv_int_alpha*)alloca(sizeof(ccv_int_alpha) * a->cols * 2);
int ch = ccv_clamp(CCV_GET_CHANNEL(a->type), 1, 4);
double scale_x = (double)a->cols / b->cols;
double scale_y = (double)a->rows / b->rows;
// double scale = 1.f / (scale_x * scale_y);
unsigned int inv_scale_256 = (int)(scale_x * scale_y * 0x10000);
int dx, dy, sx, sy, i, k;
for (dx = 0, k = 0; dx < b->cols; dx++)
{
double fsx1 = dx * scale_x, fsx2 = fsx1 + scale_x;
int sx1 = (int)(fsx1 + 1.0 - 1e-6), sx2 = (int)(fsx2);
sx1 = ccv_min(sx1, a->cols - 1);
sx2 = ccv_min(sx2, a->cols - 1);
if (sx1 > fsx1)
{
xofs[k].di = dx * ch;
xofs[k].si = (sx1 - 1) * ch;
xofs[k++].alpha = (unsigned int)((sx1 - fsx1) * 0x100);
}
for (sx = sx1; sx < sx2; sx++)
{
xofs[k].di = dx * ch;
xofs[k].si = sx * ch;
xofs[k++].alpha = 256;
}
if (fsx2 - sx2 > 1e-3)
{
xofs[k].di = dx * ch;
xofs[k].si = sx2 * ch;
xofs[k++].alpha = (unsigned int)((fsx2 - sx2) * 256);
}
}
int xofs_count = k;
unsigned int* buf = (unsigned int*)alloca(b->cols * ch * sizeof(unsigned int));
unsigned int* sum = (unsigned int*)alloca(b->cols * ch * sizeof(unsigned int));
for (dx = 0; dx < b->cols * ch; dx++)
buf[dx] = sum[dx] = 0;
dy = 0;
for (sy = 0; sy < a->rows; sy++)
{
unsigned char* a_ptr = a->data.u8 + a->step * sy;
for (k = 0; k < xofs_count; k++)
{
int dxn = xofs[k].di;
unsigned int alpha = xofs[k].alpha;
for (i = 0; i < ch; i++)
buf[dxn + i] += a_ptr[xofs[k].si + i] * alpha;
}
if ((dy + 1) * scale_y <= sy + 1 || sy == a->rows - 1)
{
unsigned int beta = (int)(ccv_max(sy + 1 - (dy + 1) * scale_y, 0.f) * 256);
unsigned int beta1 = 256 - beta;
unsigned char* b_ptr = b->data.u8 + b->step * dy;
if (beta <= 0)
{
for (dx = 0; dx < b->cols * ch; dx++)
{
b_ptr[dx] = ccv_clamp((sum[dx] + buf[dx] * 256) / inv_scale_256, 0, 255);
sum[dx] = buf[dx] = 0;
}
} else {
for (dx = 0; dx < b->cols * ch; dx++)
{
b_ptr[dx] = ccv_clamp((sum[dx] + buf[dx] * beta1) / inv_scale_256, 0, 255);
sum[dx] = buf[dx] * beta;
buf[dx] = 0;
}
}
dy++;
}
else
{
for(dx = 0; dx < b->cols * ch; dx++)
{
sum[dx] += buf[dx] * 256;
buf[dx] = 0;
}
}
}
}
ccv_array_t* ccv_bbf_detect_objects(ccv_dense_matrix_t* a, ccv_bbf_classifier_cascade_t** _cascade, int count, ccv_bbf_param_t params)
{
int hr = a->rows / ENDORSE(params.size.height);
int wr = a->cols / ENDORSE(params.size.width);
double scale = pow(2., 1. / (params.interval + 1.));
APPROX int next = params.interval + 1;
int scale_upto = (int)(log((double)ccv_min(hr, wr)) / log(scale));
ccv_dense_matrix_t** pyr = (ccv_dense_matrix_t**)alloca(ENDORSE(scale_upto + next * 2) * 4 * sizeof(ccv_dense_matrix_t*));
memset(pyr, 0, (scale_upto + next * 2) * 4 * sizeof(ccv_dense_matrix_t*));
if (ENDORSE(params.size.height != _cascade[0]->size.height || params.size.width != _cascade[0]->size.width))
ccv_resample(a, &pyr[0], 0, a->rows * ENDORSE(_cascade[0]->size.height / params.size.height), a->cols * ENDORSE(_cascade[0]->size.width / params.size.width), CCV_INTER_AREA);
else
pyr[0] = a;
APPROX int i;
int j, k, t, x, y, q;
for (i = 1; ENDORSE(i < ccv_min(params.interval + 1, scale_upto + next * 2)); i++)
ccv_resample(pyr[0], &pyr[i * 4], 0, (int)(pyr[0]->rows / pow(scale, i)), (int)(pyr[0]->cols / pow(scale, i)), CCV_INTER_AREA);
for (i = next; ENDORSE(i < scale_upto + next * 2); i++)
ccv_sample_down(pyr[i * 4 - next * 4], &pyr[i * 4], 0, 0, 0);
if (params.accurate)
for (i = next * 2; ENDORSE(i < scale_upto + next * 2); i++)
{
ccv_sample_down(pyr[i * 4 - next * 4], &pyr[i * 4 + 1], 0, 1, 0);
ccv_sample_down(pyr[i * 4 - next * 4], &pyr[i * 4 + 2], 0, 0, 1);
ccv_sample_down(pyr[i * 4 - next * 4], &pyr[i * 4 + 3], 0, 1, 1);
}
ccv_array_t* idx_seq;
ccv_array_t* seq = ccv_array_new(sizeof(ccv_comp_t), 64, 0);
ccv_array_t* seq2 = ccv_array_new(sizeof(ccv_comp_t), 64, 0);
ccv_array_t* result_seq = ccv_array_new(sizeof(ccv_comp_t), 64, 0);
/* detect in multi scale */
for (t = 0; t < count; t++)
{
ccv_bbf_classifier_cascade_t* cascade = _cascade[t];
APPROX float scale_x = (float) params.size.width / (float) cascade->size.width;
APPROX float scale_y = (float) params.size.height / (float) cascade->size.height;
ccv_array_clear(seq);
for (i = 0; ENDORSE(i < scale_upto); i++)
{
APPROX int dx[] = {0, 1, 0, 1};
APPROX int dy[] = {0, 0, 1, 1};
APPROX int i_rows = pyr[i * 4 + next * 8]->rows - ENDORSE(cascade->size.height >> 2);
APPROX int steps[] = { pyr[i * 4]->step, pyr[i * 4 + next * 4]->step, pyr[i * 4 + next * 8]->step };
APPROX int i_cols = pyr[i * 4 + next * 8]->cols - ENDORSE(cascade->size.width >> 2);
int paddings[] = { pyr[i * 4]->step * 4 - i_cols * 4,
pyr[i * 4 + next * 4]->step * 2 - i_cols * 2,
pyr[i * 4 + next * 8]->step - i_cols };
for (q = 0; q < (params.accurate ? 4 : 1); q++)
{
APPROX unsigned char* u8[] = { pyr[i * 4]->data.u8 + dx[q] * 2 + dy[q] * pyr[i * 4]->step * 2, pyr[i * 4 + next * 4]->data.u8 + dx[q] + dy[q] * pyr[i * 4 + next * 4]->step, pyr[i * 4 + next * 8 + q]->data.u8 };
for (y = 0; ENDORSE(y < i_rows); y++)
{
for (x = 0; ENDORSE(x < i_cols); x++)
{
APPROX float sum;
APPROX int flag = 1;
ccv_bbf_stage_classifier_t* classifier = cascade->stage_classifier;
for (j = 0; j < ENDORSE(cascade->count); ++j, ++classifier)
{
sum = 0;
APPROX float* alpha = classifier->alpha;
ccv_bbf_feature_t* feature = classifier->feature;
for (k = 0; k < ENDORSE(classifier->count); ++k, alpha += 2, ++feature)
sum += alpha[_ccv_run_bbf_feature(feature, ENDORSE(steps), u8)];
if (ENDORSE(sum) < ENDORSE(classifier->threshold))
{
flag = 0;
break;
}
}
if (ENDORSE(flag))
{
ccv_comp_t comp;
comp.rect = ccv_rect((int)((x * 4 + dx[q] * 2) * scale_x + 0.5), (int)((y * 4 + dy[q] * 2) * scale_y + 0.5), (int)(cascade->size.width * scale_x + 0.5), (int)(cascade->size.height * scale_y + 0.5));
comp.neighbors = 1;
comp.classification.id = t;
comp.classification.confidence = sum;
ccv_array_push(seq, &comp);
}
u8[0] += 4;
u8[1] += 2;
u8[2] += 1;
}
u8[0] += paddings[0];
u8[1] += paddings[1];
u8[2] += paddings[2];
}
}
scale_x *= scale;
scale_y *= scale;
}
/* the following code from OpenCV's haar feature implementation */
if(params.min_neighbors == 0)
{
for (i = 0; ENDORSE(i < seq->rnum); i++)
{
ccv_comp_t* comp = (ccv_comp_t*)ENDORSE(ccv_array_get(seq, i));
ccv_array_push(result_seq, comp);
}
} else {
idx_seq = 0;
ccv_array_clear(seq2);
// group retrieved rectangles in order to filter out noise
int ncomp = ccv_array_group(seq, &idx_seq, _ccv_is_equal_same_class, 0);
ccv_comp_t* comps = (ccv_comp_t*)ccmalloc((ncomp + 1) * sizeof(ccv_comp_t));
memset(comps, 0, (ncomp + 1) * sizeof(ccv_comp_t));
// count number of neighbors
for(i = 0; ENDORSE(i < seq->rnum); i++)
{
ccv_comp_t r1 = *(ccv_comp_t*)ENDORSE(ccv_array_get(seq, i));
int idx = *(int*)ENDORSE(ccv_array_get(idx_seq, i));
if (ENDORSE(comps[idx].neighbors) == 0)
comps[idx].classification.confidence = r1.classification.confidence;
++comps[idx].neighbors;
comps[idx].rect.x += r1.rect.x;
comps[idx].rect.y += r1.rect.y;
comps[idx].rect.width += r1.rect.width;
comps[idx].rect.height += r1.rect.height;
comps[idx].classification.id = r1.classification.id;
comps[idx].classification.confidence = ccv_max(comps[idx].classification.confidence, r1.classification.confidence);
}
// calculate average bounding box
for(i = 0; ENDORSE(i < ncomp); i++)
{
int n = ENDORSE(comps[i].neighbors);
if(n >= params.min_neighbors)
{
ccv_comp_t comp;
comp.rect.x = (comps[i].rect.x * 2 + n) / (2 * n);
comp.rect.y = (comps[i].rect.y * 2 + n) / (2 * n);
comp.rect.width = (comps[i].rect.width * 2 + n) / (2 * n);
comp.rect.height = (comps[i].rect.height * 2 + n) / (2 * n);
comp.neighbors = comps[i].neighbors;
comp.classification.id = comps[i].classification.id;
comp.classification.confidence = comps[i].classification.confidence;
ccv_array_push(seq2, &comp);
}
}
// filter out small face rectangles inside large face rectangles
for(i = 0; ENDORSE(i < seq2->rnum); i++)
{
ccv_comp_t r1 = *(ccv_comp_t*)ENDORSE(ccv_array_get(seq2, i));
APPROX int flag = 1;
for(j = 0; ENDORSE(j < seq2->rnum); j++)
{
ccv_comp_t r2 = *(ccv_comp_t*)ENDORSE(ccv_array_get(seq2, j));
APPROX int distance = (int)(r2.rect.width * 0.25 + 0.5);
if(ENDORSE(i != j &&
r1.classification.id == r2.classification.id &&
r1.rect.x >= r2.rect.x - distance &&
r1.rect.y >= r2.rect.y - distance &&
r1.rect.x + r1.rect.width <= r2.rect.x + r2.rect.width + distance &&
r1.rect.y + r1.rect.height <= r2.rect.y + r2.rect.height + distance &&
(r2.neighbors > ccv_max(3, r1.neighbors) || r1.neighbors < 3)))
{
flag = 0;
break;
}
}
if(ENDORSE(flag))
ccv_array_push(result_seq, &r1);
}
ccv_array_free(idx_seq);
ccfree(comps);
}
}
ccv_array_free(seq);
ccv_array_free(seq2);
ccv_array_t* result_seq2;
/* the following code from OpenCV's haar feature implementation */
if (params.flags & CCV_BBF_NO_NESTED)
{
result_seq2 = ccv_array_new(sizeof(ccv_comp_t), 64, 0);
idx_seq = 0;
// group retrieved rectangles in order to filter out noise
int ncomp = ccv_array_group(result_seq, &idx_seq, _ccv_is_equal, 0);
ccv_comp_t* comps = (ccv_comp_t*)ccmalloc((ncomp + 1) * sizeof(ccv_comp_t));
memset(comps, 0, (ncomp + 1) * sizeof(ccv_comp_t));
// count number of neighbors
for(i = 0; ENDORSE(i < result_seq->rnum); i++)
{
ccv_comp_t r1 = *(ccv_comp_t*)ENDORSE(ccv_array_get(result_seq, i));
int idx = *(int*)ENDORSE(ccv_array_get(idx_seq, i));
if (ENDORSE(comps[idx].neighbors == 0 || comps[idx].classification.confidence < r1.classification.confidence))
{
comps[idx].classification.confidence = r1.classification.confidence;
comps[idx].neighbors = 1;
comps[idx].rect = r1.rect;
comps[idx].classification.id = r1.classification.id;
}
}
// calculate average bounding box
for(i = 0; ENDORSE(i < ncomp); i++)
if(ENDORSE(comps[i].neighbors))
ccv_array_push(result_seq2, &comps[i]);
ccv_array_free(result_seq);
ccfree(comps);
} else {
result_seq2 = result_seq;
}
for (i = 1; ENDORSE(i < scale_upto + next * 2); i++)
ccv_matrix_free(pyr[i * 4]);
if (params.accurate)
for (i = next * 2; ENDORSE(i < scale_upto + next * 2); i++)
{
ccv_matrix_free(pyr[i * 4 + 1]);
ccv_matrix_free(pyr[i * 4 + 2]);
ccv_matrix_free(pyr[i * 4 + 3]);
}
if (ENDORSE(params.size.height != _cascade[0]->size.height || params.size.width != _cascade[0]->size.width))
ccv_matrix_free(pyr[0]);
return result_seq2;
}
/* sobel filter is fundamental to many other high-level algorithms,
* here includes 2 special case impl (for 1x3/3x1, 3x3) and one general impl */
void ccv_sobel(ccv_dense_matrix_t* a, ccv_dense_matrix_t** b, int type, int dx, int dy)
{
ccv_declare_derived_signature(sig, a->sig != 0, ccv_sign_with_format(64, "ccv_sobel(%d,%d)", dx, dy), a->sig, CCV_EOF_SIGN);
type = (type == 0) ? CCV_32S | CCV_GET_CHANNEL(a->type) : CCV_GET_DATA_TYPE(type) | CCV_GET_CHANNEL(a->type);
ccv_dense_matrix_t* db = *b = ccv_dense_matrix_renew(*b, a->rows, a->cols, CCV_GET_CHANNEL(a->type) | CCV_ALL_DATA_TYPE, type, sig);
ccv_object_return_if_cached(, db);
int i, j, k, c, ch = CCV_GET_CHANNEL(a->type);
unsigned char* a_ptr = a->data.u8;
unsigned char* b_ptr = db->data.u8;
if (dx == 1 || dy == 1)
{
/* special case 1: 1x3 or 3x1 window */
if (dx > dy)
{
#define for_block(_for_get, _for_set) \
for (i = 0; i < a->rows; i++) \
{ \
for (k = 0; k < ch; k++) \
_for_set(b_ptr, k, _for_get(a_ptr, ch + k, 0) - _for_get(a_ptr, k, 0), 0); \
for (j = 1; j < a->cols - 1; j++) \
for (k = 0; k < ch; k++) \
_for_set(b_ptr, j * ch + k, 2 * (_for_get(a_ptr, (j + 1) * ch + k, 0) - _for_get(a_ptr, (j - 1) * ch + k, 0)), 0); \
for (k = 0; k < ch; k++) \
_for_set(b_ptr, (a->cols - 1) * ch + k, _for_get(a_ptr, (a->cols - 1) * ch + k, 0) - _for_get(a_ptr, (a->cols - 2) * ch + k, 0), 0); \
b_ptr += db->step; \
a_ptr += a->step; \
}
ccv_matrix_getter(a->type, ccv_matrix_setter, db->type, for_block);
#undef for_block
} else {
#define for_block(_for_get, _for_set) \
for (j = 0; j < a->cols; j++) \
for (k = 0; k < ch; k++) \
_for_set(b_ptr, j * ch + k, _for_get(a_ptr + a->step, j * ch + k, 0) - _for_get(a_ptr, j * ch + k, 0), 0); \
a_ptr += a->step; \
b_ptr += db->step; \
for (i = 1; i < a->rows - 1; i++) \
{ \
for (j = 0; j < a->cols; j++) \
for (k = 0; k < ch; k++) \
_for_set(b_ptr, j * ch + k, 2 * (_for_get(a_ptr + a->step, j * ch + k, 0) - _for_get(a_ptr - a->step, j * ch + k, 0)), 0); \
a_ptr += a->step; \
b_ptr += db->step; \
} \
for (j = 0; j < a->cols; j++) \
for (k = 0; k < ch; k++) \
_for_set(b_ptr, j * ch + k, _for_get(a_ptr, j * ch + k, 0) - _for_get(a_ptr - a->step, j * ch + k, 0), 0);
ccv_matrix_getter(a->type, ccv_matrix_setter, db->type, for_block);
#undef for_block
}
} else if (dx > 3 || dy > 3) {
/* general case: in this case, I will generate a separable filter, and do the convolution */
int fsz = ccv_max(dx, dy);
assert(fsz % 2 == 1);
int hfz = fsz / 2;
unsigned char* df = (unsigned char*)alloca(sizeof(double) * fsz);
unsigned char* gf = (unsigned char*)alloca(sizeof(double) * fsz);
/* the sigma calculation is linear derviation of 3x3 - 0.85, 5x5 - 1.32 */
double sigma = ((fsz - 1) / 2) * 0.47 + 0.38;
double sigma2 = (2.0 * sigma * sigma);
/* 2.5 is the factor to make the kernel "visible" in integer setting */
double psigma3 = 2.5 / sqrt(sqrt(2 * CCV_PI) * sigma * sigma * sigma);
for (i = 0; i < fsz; i++)
{
((double*)df)[i] = (i - hfz) * exp(-((i - hfz) * (i - hfz)) / sigma2) * psigma3;
((double*)gf)[i] = exp(-((i - hfz) * (i - hfz)) / sigma2) * psigma3;
}
if (db->type & CCV_32S)
{
for (i = 0; i < fsz; i++)
{
// df could be negative, thus, (int)(x + 0.5) shortcut will not work
((int*)df)[i] = (int)round(((double*)df)[i] * 256.0);
((int*)gf)[i] = (int)(((double*)gf)[i] * 256.0 + 0.5);
}
} else {
for (i = 0; i < fsz; i++)
{
ccv_set_value(db->type, df, i, ((double*)df)[i], 0);
ccv_set_value(db->type, gf, i, ((double*)gf)[i], 0);
}
}
if (dx < dy)
{
unsigned char* tf = df;
df = gf;
gf = tf;
}
unsigned char* buf = (unsigned char*)alloca(sizeof(double) * ch * (fsz + ccv_max(a->rows, a->cols)));
#define for_block(_for_get, _for_type_b, _for_set_b, _for_get_b) \
for (i = 0; i < a->rows; i++) \
{ \
for (j = 0; j < hfz; j++) \
for (k = 0; k < ch; k++) \
_for_set_b(buf, j * ch + k, _for_get(a_ptr, k, 0), 0); \
for (j = 0; j < a->cols; j++) \
for (k = 0; k < ch; k++) \
_for_set_b(buf, (j + hfz) * ch + k, _for_get(a_ptr, j * ch + k, 0), 0); \
for (j = a->cols; j < a->cols + hfz; j++) \
for (k = 0; k < ch; k++) \
_for_set_b(buf, (j + hfz) * ch + k, _for_get(a_ptr, (a->cols - 1) * ch + k, 0), 0); \
for (j = 0; j < a->cols; j++) \
{ \
for (c = 0; c < ch; c++) \
{ \
_for_type_b sum = 0; \
for (k = 0; k < fsz; k++) \
sum += _for_get_b(buf, (j + k) * ch + c, 0) * _for_get_b(df, k, 0); \
_for_set_b(b_ptr, j * ch + c, sum, 8); \
} \
} \
a_ptr += a->step; \
b_ptr += db->step; \
} \
b_ptr = db->data.u8; \
for (i = 0; i < a->cols; i++) \
{ \
for (j = 0; j < hfz; j++) \
for (k = 0; k < ch; k++) \
_for_set_b(buf, j * ch + k, _for_get_b(b_ptr, i * ch + k, 0), 0); \
for (j = 0; j < a->rows; j++) \
for (k = 0; k < ch; k++) \
_for_set_b(buf, (j + hfz) * ch + k, _for_get_b(b_ptr + j * db->step, i * ch + k, 0), 0); \
for (j = a->rows; j < a->rows + hfz; j++) \
for (k = 0; k < ch; k++) \
_for_set_b(buf, (j + hfz) * ch + k, _for_get_b(b_ptr + (a->rows - 1) * db->step, i * ch + k, 0), 0); \
for (j = 0; j < a->rows; j++) \
{ \
for (c = 0; c < ch; c++) \
{ \
_for_type_b sum = 0; \
for (k = 0; k < fsz; k++) \
sum += _for_get_b(buf, (j + k) * ch + c, 0) * _for_get_b(gf, k, 0); \
_for_set_b(b_ptr + j * db->step, i * ch + c, sum, 8); \
} \
} \
}
ccv_matrix_getter(a->type, ccv_matrix_typeof_setter_getter, db->type, for_block);
#undef for_block
} else {
/* special case 2: 3x3 window, corresponding sigma = 0.85 */
unsigned char* buf = (unsigned char*)alloca(db->step);
if (dx > dy)
{
#define for_block(_for_get, _for_set_b, _for_get_b) \
for (j = 0; j < a->cols; j++) \
for (k = 0; k < ch; k++) \
_for_set_b(b_ptr, j * ch + k, _for_get(a_ptr + a->step, j * ch + k, 0) + 3 * _for_get(a_ptr, j * ch + k, 0), 0); \
a_ptr += a->step; \
b_ptr += db->step; \
for (i = 1; i < a->rows - 1; i++) \
{ \
for (j = 0; j < a->cols; j++) \
for (k = 0; k < ch; k++) \
_for_set_b(b_ptr, j * ch + k, _for_get(a_ptr + a->step, j * ch + k, 0) + 2 * _for_get(a_ptr, j * ch + k, 0) + _for_get(a_ptr - a->step, j * ch + k, 0), 0); \
a_ptr += a->step; \
b_ptr += db->step; \
} \
for (j = 0; j < a->cols; j++) \
for (k = 0; k < ch; k++) \
_for_set_b(b_ptr, j * ch + k, 3 * _for_get(a_ptr, j * ch + k, 0) + _for_get(a_ptr - a->step, j * ch + k, 0), 0); \
b_ptr = db->data.u8; \
for (i = 0; i < a->rows; i++) \
{ \
for (k = 0; k < ch; k++) \
_for_set_b(buf, k, _for_get_b(b_ptr, ch + k, 0) - _for_get_b(b_ptr, k, 0), 0); \
for (j = 1; j < a->cols - 1; j++) \
for (k = 0; k < ch; k++) \
_for_set_b(buf, j * ch + k, _for_get_b(b_ptr, (j + 1) * ch + k, 0) - _for_get_b(b_ptr, (j - 1) * ch + k, 0), 0); \
for (k = 0; k < ch; k++) \
_for_set_b(buf, (a->cols - 1) * ch + k, _for_get_b(b_ptr, (a->cols - 1) * ch + k, 0) - _for_get_b(b_ptr, (a->cols - 2) * ch + k, 0), 0); \
memcpy(b_ptr, buf, db->step); \
b_ptr += db->step; \
}
ccv_matrix_getter(a->type, ccv_matrix_setter_getter, db->type, for_block);
#undef for_block
} else {
#define for_block(_for_get, _for_set_b, _for_get_b) \
for (j = 0; j < a->cols; j++) \
for (k = 0; k < ch; k++) \
_for_set_b(b_ptr, j * ch + k, _for_get(a_ptr + a->step, j * ch + k, 0) - _for_get(a_ptr, j * ch + k, 0), 0); \
a_ptr += a->step; \
b_ptr += db->step; \
for (i = 1; i < a->rows - 1; i++) \
{ \
for (j = 0; j < a->cols; j++) \
for (k = 0; k < ch; k++) \
_for_set_b(b_ptr, j * ch + k, _for_get(a_ptr + a->step, j * ch + k, 0) - _for_get(a_ptr - a->step, j * ch + k, 0), 0); \
a_ptr += a->step; \
b_ptr += db->step; \
} \
for (j = 0; j < a->cols; j++) \
for (k = 0; k < ch; k++) \
_for_set_b(b_ptr, j * ch + k, _for_get(a_ptr, j * ch + k, 0) - _for_get(a_ptr - a->step, j * ch + k, 0), 0); \
b_ptr = db->data.u8; \
for (i = 0; i < a->rows; i++) \
{ \
for (k = 0; k < ch; k++) \
_for_set_b(buf, k, _for_get_b(b_ptr, ch + k, 0) + 3 * _for_get_b(b_ptr, k, 0), 0); \
for (j = 1; j < a->cols - 1; j++) \
for (k = 0; k < ch; k++) \
_for_set_b(buf, j * ch + k, _for_get_b(b_ptr, (j + 1) * ch + k, 0) + 2 * _for_get_b(b_ptr, j * ch + k, 0) + _for_get_b(b_ptr, (j - 1) * ch + k, 0), 0); \
for (k = 0; k < ch; k++) \
_for_set_b(buf, (a->cols - 1) * ch + k, _for_get_b(b_ptr, (a->cols - 2) * ch + k, 0) + 3 * _for_get_b(b_ptr, (a->cols - 1) * ch + k, 0), 0); \
memcpy(b_ptr, buf, db->step); \
b_ptr += db->step; \
}
ccv_matrix_getter(a->type, ccv_matrix_setter_getter, db->type, for_block);
#undef for_block
}
}
}
void ccv_blur(ccv_dense_matrix_t* a, ccv_dense_matrix_t** b, int type, double sigma)
{
ccv_declare_derived_signature(sig, a->sig != 0, ccv_sign_with_format(64, "ccv_blur(%la)", sigma), a->sig, CCV_EOF_SIGN);
type = (type == 0) ? CCV_GET_DATA_TYPE(a->type) | CCV_GET_CHANNEL(a->type) : CCV_GET_DATA_TYPE(type) | CCV_GET_CHANNEL(a->type);
ccv_dense_matrix_t* db = *b = ccv_dense_matrix_renew(*b, a->rows, a->cols, CCV_ALL_DATA_TYPE | CCV_GET_CHANNEL(a->type), type, sig);
ccv_object_return_if_cached(, db);
int fsz = ccv_max(1, (int)(4.0 * sigma + 1.0 - 1e-8)) * 2 + 1;
int hfz = fsz / 2;
unsigned char* buf = (unsigned char*)alloca(sizeof(double) * ccv_max(fsz + a->rows, (fsz + a->cols) * CCV_GET_CHANNEL(a->type)));
unsigned char* filter = (unsigned char*)alloca(sizeof(double) * fsz);
double tw = 0;
int i, j, k, ch = CCV_GET_CHANNEL(a->type);
for (i = 0; i < fsz; i++)
tw += ((double*)filter)[i] = exp(-((i - hfz) * (i - hfz)) / (2.0 * sigma * sigma));
int no_8u_type = (db->type & CCV_8U) ? CCV_32S : db->type;
if (no_8u_type & CCV_32S)
{
tw = 256.0 / tw;
for (i = 0; i < fsz; i++)
((int*)filter)[i] = (int)(((double*)filter)[i] * tw + 0.5);
} else {
tw = 1.0 / tw;
for (i = 0; i < fsz; i++)
ccv_set_value(db->type, filter, i, ((double*)filter)[i] * tw, 0);
}
/* horizontal */
unsigned char* a_ptr = a->data.u8;
unsigned char* b_ptr = db->data.u8;
#define for_block(_for_type, _for_set_b, _for_get_b, _for_set_a, _for_get_a) \
for (i = 0; i < a->rows; i++) \
{ \
for (j = 0; j < hfz; j++) \
for (k = 0; k < ch; k++) \
_for_set_b(buf, j * ch + k, _for_get_a(a_ptr, k, 0), 0); \
for (j = 0; j < a->cols * ch; j++) \
_for_set_b(buf, j + hfz * ch, _for_get_a(a_ptr, j, 0), 0); \
for (j = a->cols; j < hfz + a->cols; j++) \
for (k = 0; k < ch; k++) \
_for_set_b(buf, j * ch + hfz * ch + k, _for_get_a(a_ptr, (a->cols - 1) * ch + k, 0), 0); \
for (j = 0; j < a->cols * ch; j++) \
{ \
_for_type sum = 0; \
for (k = 0; k < fsz; k++) \
sum += _for_get_b(buf, k * ch + j, 0) * _for_get_b(filter, k, 0); \
_for_set_b(buf, j, sum, 8); \
} \
for (j = 0; j < a->cols * ch; j++) \
_for_set_a(b_ptr, j, _for_get_b(buf, j, 0), 0); \
a_ptr += a->step; \
b_ptr += db->step; \
}
ccv_matrix_typeof_setter_getter(no_8u_type, ccv_matrix_setter, db->type, ccv_matrix_getter, a->type, for_block);
#undef for_block
/* vertical */
b_ptr = db->data.u8;
#define for_block(_for_type, _for_set_b, _for_get_b, _for_set_a, _for_get_a) \
for (i = 0; i < a->cols * ch; i++) \
{ \
for (j = 0; j < hfz; j++) \
_for_set_b(buf, j, _for_get_a(b_ptr, i, 0), 0); \
for (j = 0; j < a->rows; j++) \
_for_set_b(buf, j + hfz, _for_get_a(b_ptr + j * db->step, i, 0), 0); \
for (j = a->rows; j < hfz + a->rows; j++) \
_for_set_b(buf, j + hfz, _for_get_a(b_ptr + (a->rows - 1) * db->step, i, 0), 0); \
for (j = 0; j < a->rows; j++) \
{ \
_for_type sum = 0; \
for (k = 0; k < fsz; k++) \
sum += _for_get_b(buf, k + j, 0) * _for_get_b(filter, k, 0); \
_for_set_b(buf, j, sum, 8); \
} \
for (j = 0; j < a->rows; j++) \
_for_set_a(b_ptr + j * db->step, i, _for_get_b(buf, j, 0), 0); \
}
ccv_matrix_typeof_setter_getter(no_8u_type, ccv_matrix_setter_getter, db->type, for_block);
#undef for_block
}
ccv_decimal_point_t ccv_perspective_transform_apply(ccv_decimal_point_t point, ccv_size_t size, float m00, float m01, float m02, float m10, float m11, float m12, float m20, float m21, float m22)
{
m00 *= 1.0 / ccv_max(size.width, size.height);
m01 *= 1.0 / ccv_max(size.width, size.height);
m02 *= 1.0 / ccv_max(size.width, size.height);
m10 *= 1.0 / ccv_max(size.width, size.height);
m11 *= 1.0 / ccv_max(size.width, size.height);
m12 *= 1.0 / ccv_max(size.width, size.height);
m20 *= 1.0 / (ccv_max(size.width, size.height) * ccv_max(size.width, size.height));
m21 *= 1.0 / (ccv_max(size.width, size.height) * ccv_max(size.width, size.height));
m22 *= 1.0 / ccv_max(size.width, size.height);
point.x -= size.width * 0.5;
point.y -= size.height * 0.5;
float wz = 1.0 / (point.x * m20 + point.y * m21 + m22);
float wx = size.width * 0.5 + (point.x * m00 + point.y * m01 + m02) * wz;
float wy = size.height * 0.5 + (point.x * m10 + point.y * m11 + m12) * wz;
return ccv_decimal_point(wx, wy);
}
// this method is a merely baseline implementation and has no optimization effort ever put into it, if at all
void ccv_perspective_transform(ccv_dense_matrix_t* a, ccv_dense_matrix_t** b, int type, float m00, float m01, float m02, float m10, float m11, float m12, float m20, float m21, float m22)
{
ccv_declare_derived_signature(sig, a->sig != 0, ccv_sign_with_format(64, "ccv_perspective_transform(%a,%a,%a,%a,%a,%a,%a,%a,%a)", m00, m01, m02, m10, m11, m12, m20, m21, m22), a->sig, CCV_EOF_SIGN);
type = (type == 0) ? CCV_GET_DATA_TYPE(a->type) | CCV_GET_CHANNEL(a->type) : CCV_GET_DATA_TYPE(type) | CCV_GET_CHANNEL(a->type);
ccv_dense_matrix_t* db = *b = ccv_dense_matrix_renew(*b, a->rows, a->cols, CCV_ALL_DATA_TYPE | CCV_GET_CHANNEL(a->type), type, sig);
ccv_object_return_if_cached(, db);
// with default of bilinear interpolation
int i, j, k, ch = CCV_GET_CHANNEL(a->type);
unsigned char* a_ptr = a->data.u8;
unsigned char* b_ptr = db->data.u8;
// assume field of view is 60, modify the matrix value to reflect that
// (basically, apply x / ccv_max(a->rows, a->cols), y / ccv_max(a->rows, a->cols) before hand
m00 *= 1.0 / ccv_max(a->rows, a->cols);
m01 *= 1.0 / ccv_max(a->rows, a->cols);
m02 *= 1.0 / ccv_max(a->rows, a->cols);
m10 *= 1.0 / ccv_max(a->rows, a->cols);
m11 *= 1.0 / ccv_max(a->rows, a->cols);
m12 *= 1.0 / ccv_max(a->rows, a->cols);
m20 *= 1.0 / (ccv_max(a->rows, a->cols) * ccv_max(a->rows, a->cols));
m21 *= 1.0 / (ccv_max(a->rows, a->cols) * ccv_max(a->rows, a->cols));
m22 *= 1.0 / ccv_max(a->rows, a->cols);
#define for_block(_for_set, _for_get) \
for (i = 0; i < db->rows; i++) \
{ \
float cy = i - db->rows * 0.5; \
float crx = cy * m01 + m02; \
float cry = cy * m11 + m12; \
float crz = cy * m21 + m22; \
for (j = 0; j < db->cols; j++) \
{ \
float cx = j - db->cols * 0.5; \
float wz = 1.0 / (cx * m20 + crz); \
float wx = a->cols * 0.5 + (cx * m00 + crx) * wz; \
float wy = a->rows * 0.5 + (cx * m10 + cry) * wz; \
int iwx = (int)wx; \
int iwy = (int)wy; \
wx = wx - iwx; \
wy = wy - iwy; \
if (iwx >= 0 && iwx < a->cols - 1 && iwy >= 0 && iwy < a->rows - 1) \
for (k = 0; k < ch; k++) \
_for_set(b_ptr, j * ch + k, _for_get(a_ptr + iwy * a->step, iwx * ch + k, 0) * (1 - wx) * (1 - wy) + \
_for_get(a_ptr + iwy * a->step, iwx * ch + ch + k, 0) * wx * (1 - wy) + \
_for_get(a_ptr + iwy * a->step + a->step, iwx * ch + k, 0) * (1 - wx) * wy + \
_for_get(a_ptr + iwy * a->step + a->step, iwx * ch + ch + k, 0) * wx * wy, 0); \
else \
for (k = 0; k < ch; k++) \
_for_set(b_ptr, j * ch + k, 0, 0); \
} \
b_ptr += db->step; \
}
ccv_matrix_setter(db->type, ccv_matrix_getter, a->type, for_block);
#undef for_block
}
/* this code is a rewrite from OpenCV's legendary Lucas-Kanade optical flow implementation */
void ccv_optical_flow_lucas_kanade(ccv_dense_matrix_t* a, ccv_dense_matrix_t* b, ccv_array_t* point_a, ccv_array_t** point_b, ccv_size_t win_size, int level, double min_eigen)
{
assert(a && b && a->rows == b->rows && a->cols == b->cols);
assert(CCV_GET_CHANNEL(a->type) == CCV_GET_CHANNEL(b->type) && CCV_GET_DATA_TYPE(a->type) == CCV_GET_DATA_TYPE(b->type));
assert(CCV_GET_CHANNEL(a->type) == 1);
assert(CCV_GET_DATA_TYPE(a->type) == CCV_8U);
assert(point_a->rnum > 0);
level = ccv_clamp(level + 1, 1, (int)(log((double)ccv_min(a->rows, a->cols) / ccv_max(win_size.width * 2, win_size.height * 2)) / log(2.0) + 0.5));
ccv_declare_derived_signature(sig, a->sig != 0 && b->sig != 0 && point_a->sig != 0, ccv_sign_with_format(128, "ccv_optical_flow_lucas_kanade(%d,%d,%d,%la)", win_size.width, win_size.height, level, min_eigen), a->sig, b->sig, point_a->sig, CCV_EOF_SIGN);
ccv_array_t* seq = *point_b = ccv_array_new(sizeof(ccv_decimal_point_with_status_t), point_a->rnum, sig);
ccv_object_return_if_cached(, seq);
seq->rnum = point_a->rnum;
ccv_dense_matrix_t** pyr_a = (ccv_dense_matrix_t**)malloc(sizeof(ccv_dense_matrix_t*) * level);
ccv_dense_matrix_t** pyr_a_dx = (ccv_dense_matrix_t**)malloc(sizeof(ccv_dense_matrix_t*) * level);
ccv_dense_matrix_t** pyr_a_dy = (ccv_dense_matrix_t**)malloc(sizeof(ccv_dense_matrix_t*) * level);
ccv_dense_matrix_t** pyr_b = (ccv_dense_matrix_t**)malloc(sizeof(ccv_dense_matrix_t*) * level);
int i, j, t, x, y;
/* generating image pyramid */
pyr_a[0] = a;
pyr_a_dx[0] = pyr_a_dy[0] = 0;
ccv_sobel(pyr_a[0], &pyr_a_dx[0], 0, 3, 0);
ccv_sobel(pyr_a[0], &pyr_a_dy[0], 0, 0, 3);
pyr_b[0] = b;
for (i = 1; i < level; i++)
{
pyr_a[i] = pyr_a_dx[i] = pyr_a_dy[i] = pyr_b[i] = 0;
ccv_sample_down(pyr_a[i - 1], &pyr_a[i], 0, 0, 0);
ccv_sobel(pyr_a[i], &pyr_a_dx[i], 0, 3, 0);
ccv_sobel(pyr_a[i], &pyr_a_dy[i], 0, 0, 3);
ccv_sample_down(pyr_b[i - 1], &pyr_b[i], 0, 0, 0);
}
int* wi = (int*)malloc(sizeof(int) * win_size.width * win_size.height);
int* widx = (int*)malloc(sizeof(int) * win_size.width * win_size.height);
int* widy = (int*)malloc(sizeof(int) * win_size.width * win_size.height);
ccv_decimal_point_t half_win = ccv_decimal_point((win_size.width - 1) * 0.5f, (win_size.height - 1) * 0.5f);
const int W_BITS14 = 14, W_BITS7 = 7, W_BITS9 = 9;
const float FLT_SCALE = 1.0f / (1 << 25);
// clean up status to 1
for (i = 0; i < point_a->rnum; i++)
{
ccv_decimal_point_with_status_t* point_with_status = (ccv_decimal_point_with_status_t*)ccv_array_get(seq, i);
point_with_status->status = 1;
}
int prev_rows, prev_cols;
for (t = level - 1; t >= 0; t--)
{
ccv_dense_matrix_t* a = pyr_a[t];
ccv_dense_matrix_t* adx = pyr_a_dx[t];
ccv_dense_matrix_t* ady = pyr_a_dy[t];
assert(CCV_GET_DATA_TYPE(adx->type) == CCV_32S);
assert(CCV_GET_DATA_TYPE(ady->type) == CCV_32S);
ccv_dense_matrix_t* b = pyr_b[t];
for (i = 0; i < point_a->rnum; i++)
{
ccv_decimal_point_t prev_point = *(ccv_decimal_point_t*)ccv_array_get(point_a, i);
ccv_decimal_point_with_status_t* point_with_status = (ccv_decimal_point_with_status_t*)ccv_array_get(seq, i);
prev_point.x = prev_point.x / (float)(1 << t);
prev_point.y = prev_point.y / (float)(1 << t);
ccv_decimal_point_t next_point;
if (t == level - 1)
next_point = prev_point;
else {
next_point.x = point_with_status->point.x * 2 + (a->cols - prev_cols * 2) * 0.5;
next_point.y = point_with_status->point.y * 2 + (a->rows - prev_rows * 2) * 0.5;
}
point_with_status->point = next_point;
prev_point.x -= half_win.x;
prev_point.y -= half_win.y;
ccv_point_t iprev_point = ccv_point((int)prev_point.x, (int)prev_point.y);
if (iprev_point.x < 0 || iprev_point.x >= a->cols - win_size.width - 1 ||
iprev_point.y < 0 || iprev_point.y >= a->rows - win_size.height - 1)
{
if (t == 0)
point_with_status->status = 0;
continue;
}
float xd = prev_point.x - iprev_point.x;
float yd = prev_point.y - iprev_point.y;
int iw00 = (int)((1 - xd) * (1 - yd) * (1 << W_BITS14) + 0.5);
int iw01 = (int)(xd * (1 - yd) * (1 << W_BITS14) + 0.5);
int iw10 = (int)((1 - xd) * yd * (1 << W_BITS14) + 0.5);
int iw11 = (1 << W_BITS14) - iw00 - iw01 - iw10;
float a11 = 0, a12 = 0, a22 = 0;
unsigned char* a_ptr = (unsigned char*)ccv_get_dense_matrix_cell_by(CCV_C1 | CCV_8U, a, iprev_point.y, iprev_point.x, 0);
int* adx_ptr = (int*)ccv_get_dense_matrix_cell_by(CCV_C1 | CCV_32S, adx, iprev_point.y, iprev_point.x, 0);
int* ady_ptr = (int*)ccv_get_dense_matrix_cell_by(CCV_C1 | CCV_32S, ady, iprev_point.y, iprev_point.x, 0);
int* wi_ptr = wi;
int* widx_ptr = widx;
int* widy_ptr = widy;
for (y = 0; y < win_size.height; y++)
{
for (x = 0; x < win_size.width; x++)
{
wi_ptr[x] = ccv_descale(a_ptr[x] * iw00 + a_ptr[x + 1] * iw01 + a_ptr[x + a->step] * iw10 + a_ptr[x + a->step + 1] * iw11, W_BITS7);
// because we use 3x3 sobel, which scaled derivative up by 4
widx_ptr[x] = ccv_descale(adx_ptr[x] * iw00 + adx_ptr[x + 1] * iw01 + adx_ptr[x + adx->cols] * iw10 + adx_ptr[x + adx->cols + 1] * iw11, W_BITS9);
widy_ptr[x] = ccv_descale(ady_ptr[x] * iw00 + ady_ptr[x + 1] * iw01 + ady_ptr[x + ady->cols] + iw10 + ady_ptr[x + ady->cols + 1] * iw11, W_BITS9);
a11 += (float)(widx_ptr[x] * widx_ptr[x]);
a12 += (float)(widx_ptr[x] * widy_ptr[x]);
a22 += (float)(widy_ptr[x] * widy_ptr[x]);
}
a_ptr += a->step;
adx_ptr += adx->cols;
ady_ptr += ady->cols;
wi_ptr += win_size.width;
widx_ptr += win_size.width;
widy_ptr += win_size.width;
}
a11 *= FLT_SCALE;
a12 *= FLT_SCALE;
a22 *= FLT_SCALE;
float D = a11 * a22 - a12 * a12;
float eigen = (a22 + a11 - sqrtf((a11 - a22) * (a11 - a22) + 4.0f * a12 * a12)) / (2 * win_size.width * win_size.height);
if (eigen < min_eigen || D < FLT_EPSILON)
{
if (t == 0)
point_with_status->status = 0;
continue;
}
D = 1.0f / D;
next_point.x -= half_win.x;
next_point.y -= half_win.y;
ccv_decimal_point_t prev_delta;
for (j = 0; j < LK_MAX_ITER; j++)
{
ccv_point_t inext_point = ccv_point((int)next_point.x, (int)next_point.y);
if (inext_point.x < 0 || inext_point.x >= a->cols - win_size.width - 1 ||
inext_point.y < 0 || inext_point.y >= a->rows - win_size.height - 1)
break;
float xd = next_point.x - inext_point.x;
float yd = next_point.y - inext_point.y;
int iw00 = (int)((1 - xd) * (1 - yd) * (1 << W_BITS14) + 0.5);
int iw01 = (int)(xd * (1 - yd) * (1 << W_BITS14) + 0.5);
int iw10 = (int)((1 - xd) * yd * (1 << W_BITS14) + 0.5);
int iw11 = (1 << W_BITS14) - iw00 - iw01 - iw10;
float b1 = 0, b2 = 0;
unsigned char* b_ptr = (unsigned char*)ccv_get_dense_matrix_cell_by(CCV_C1 | CCV_8U, b, inext_point.y, inext_point.x, 0);
int* wi_ptr = wi;
int* widx_ptr = widx;
int* widy_ptr = widy;
for (y = 0; y < win_size.height; y++)
{
for (x = 0; x < win_size.width; x++)
{
int diff = ccv_descale(b_ptr[x] * iw00 + b_ptr[x + 1] * iw01 + b_ptr[x + b->step] * iw10 + b_ptr[x + b->step + 1] * iw11, W_BITS7) - wi_ptr[x];
b1 += (float)(diff * widx_ptr[x]);
b2 += (float)(diff * widy_ptr[x]);
}
b_ptr += b->step;
wi_ptr += win_size.width;
widx_ptr += win_size.width;
widy_ptr += win_size.width;
}
b1 *= FLT_SCALE;
b2 *= FLT_SCALE;
ccv_decimal_point_t delta = ccv_decimal_point((a12 * b2 - a22 * b1) * D, (a12 * b1 - a11 * b2) * D);
next_point.x += delta.x;
next_point.y += delta.y;
if (delta.x * delta.x + delta.y * delta.y < LK_EPSILON)
break;
if (j > 0 && fabs(prev_delta.x - delta.x) < 0.01 && fabs(prev_delta.y - delta.y) < 0.01)
{
next_point.x -= delta.x * 0.5;
next_point.y -= delta.y * 0.5;
break;
}
prev_delta = delta;
}
ccv_point_t inext_point = ccv_point((int)next_point.x, (int)next_point.y);
if (inext_point.x < 0 || inext_point.x >= a->cols - win_size.width - 1 ||
inext_point.y < 0 || inext_point.y >= a->rows - win_size.height - 1)
point_with_status->status = 0;
else {
point_with_status->point.x = next_point.x + half_win.x;
point_with_status->point.y = next_point.y + half_win.y;
}
}
prev_rows = a->rows;
prev_cols = a->cols;
ccv_matrix_free(adx);
ccv_matrix_free(ady);
if (t > 0)
{
ccv_matrix_free(a);
ccv_matrix_free(b);
}
}
free(widy);
free(widx);
free(wi);
free(pyr_b);
free(pyr_a_dy);
free(pyr_a_dx);
free(pyr_a);
}
// compute harmonic mean of precision / recall of swt
static void _ccv_evaluate_wolf(ccv_array_t* words, ccv_array_t* truth, ccv_swt_param_t params, double* precision, double* recall)
{
if (words->rnum == 0 || truth->rnum == 0)
return;
int j, k;
double total_recall = 0, total_precision = 0;
int* cG = (int*)ccmalloc(sizeof(int) * truth->rnum);
int* cD = (int*)ccmalloc(sizeof(int) * words->rnum);
memset(cG, 0, sizeof(int) * truth->rnum);
memset(cD, 0, sizeof(int) * words->rnum);
double* mG = (double*)ccmalloc(sizeof(double) * truth->rnum * words->rnum);
double* mD = (double*)ccmalloc(sizeof(double) * truth->rnum * words->rnum);
memset(mG, 0, sizeof(double) * truth->rnum * words->rnum);
memset(mD, 0, sizeof(double) * truth->rnum * words->rnum);
for (j = 0; j < truth->rnum; j++)
{
ccv_rect_t* rect = (ccv_rect_t*)ccv_array_get(truth, j);
for (k = 0; k < words->rnum; k++)
{
ccv_rect_t* target = (ccv_rect_t*)ccv_array_get(words, k);
int match = ccv_max(ccv_min(target->x + target->width, rect->x + rect->width) - ccv_max(target->x, rect->x), 0) * ccv_max(ccv_min(target->y + target->height, rect->y + rect->height) - ccv_max(target->y, rect->y), 0);
if (match > 0)
{
mG[j * words->rnum + k] = (double)match / (double)(rect->width * rect->height);
mD[k * truth->rnum + j] = (double)match / (double)(target->width * target->height);
++cG[j];
++cD[k];
}
}
}
unsigned char* tG = (unsigned char*)ccmalloc(truth->rnum);
unsigned char* tD = (unsigned char*)ccmalloc(words->rnum);
memset(tG, 0, truth->rnum);
memset(tD, 0, words->rnum);
// one to one match
for (j = 0; j < truth->rnum; j++)
{
if (cG[j] != 1)
continue;
ccv_rect_t* rect = (ccv_rect_t*)ccv_array_get(truth, j);
for (k = 0; k < words->rnum; k++)
{
if (cD[k] != 1)
continue;
ccv_rect_t* target = (ccv_rect_t*)ccv_array_get(words, k);
if (mG[j * words->rnum + k] >= one_g && mD[k * truth->rnum + j] >= one_d)
{
double dx = (target->x + target->width * 0.5) - (rect->x + rect->width * 0.5);
double dy = (target->y + target->height * 0.5) - (rect->y + rect->height * 0.5);
double d = sqrt(dx * dx + dy * dy) * 2.0 / (sqrt(target->width * target->width + target->height * target->height) + sqrt(rect->width * rect->width + rect->height * rect->height));
if (d < center_diff_thr)
{
total_recall += 1.0;
total_precision += 1.0;
assert(tG[j] == 0);
assert(tD[k] == 0);
tG[j] = tD[k] = 1;
}
}
}
}
int* many = (int*)ccmalloc(sizeof(int) * ccv_max(words->rnum, truth->rnum));
// one to many match, starts with ground truth
for (j = 0; j < truth->rnum; j++)
{
if (tG[j] || cG[j] <= 1)
continue;
double one_sum = 0;
int no_many = 0;
for (k = 0; k < words->rnum; k++)
{
if (tD[k])
continue;
double many_single = mD[k * truth->rnum + j];
if (many_single >= one_d)
{
one_sum += mG[j * words->rnum + k];
many[no_many] = k;
++no_many;
}
}
if (no_many == 1)
{
// degrade to one to one match
if (mG[j * words->rnum + many[0]] >= one_g && mD[many[0] * truth->rnum + j] >= one_d)
{
total_recall += 1.0;
total_precision += 1.0;
tG[j] = tD[many[0]] = 1;
}
} else if (one_sum >= one_g) {
for (k = 0; k < no_many; k++)
tD[many[k]] = 1;
total_recall += om_one;
total_precision += om_one / (1 + log(no_many));
}
}
// one to many match, with estimate
for (k = 0; k < words->rnum; k++)
{
if (tD[k] || cD[k] <= 1)
continue;
double one_sum = 0;
int no_many = 0;
for (j = 0; j < truth->rnum; j++)
{
if (tG[j])
continue;
double many_single = mG[j * words->rnum + k];
if (many_single >= one_g)
{
one_sum += mD[k * truth->rnum + j];
many[no_many] = j;
++no_many;
}
}
if (no_many == 1)
{
// degrade to one to one match
if (mG[many[0] * words->rnum + k] >= one_g && mD[k * truth->rnum + many[0]] >= one_d)
{
total_recall += 1.0;
total_precision += 1.0;
tG[many[0]] = tD[k] = 1;
}
} else if (one_sum >= one_g) {
for (j = 0; j < no_many; j++)
tG[many[j]] = 1;
total_recall += om_one / (1 + log(no_many));
total_precision += om_one;
}
}
ccfree(many);
ccfree(tG);
ccfree(tD);
ccfree(cG);
ccfree(cD);
ccfree(mG);
ccfree(mD);
assert(total_precision < words->rnum + 0.1);
assert(total_recall < truth->rnum + 0.1);
if (precision)
*precision = total_precision;
if (recall)
*recall = total_recall;
}
