void
fastnc_detector<A>::new_particles(F& feature, PS& pset, const cpu&)
{
SCOPE_PROF(fastnc_new_particles_detector);
memset(new_points_, 0);
typename PS::kernel_type pset_ = pset;
mt_apply2d(sizeof(i_float1), saliency_.domain() - border(8),
[this, &feature, &pset_] (i_int2 p)
{
if (pset_.has(p)) return;
if (saliency_(p) == 0) return;
for (int i = 0; i < 8; i++)
{
i_int2 n(p + i_int2(c8_h[i]));
if (saliency_(p) < saliency_(n) || pset_.has(n))
return;
}
new_points_(p) = p;
}, cpu());
st_apply2d(sizeof(i_float1), saliency_.domain() - border(8),
[this, &feature, &pset] (i_int2 p)
{
if (new_points_(p) != i_short2(0,0)) pset.add(p, feature(p));
}, cpu());
}
inline __host__ __device__
gl8u compute_contrast(i_short2 p, const I& input)
{
const int d = 5;
return ::abs(int(input(p + i_int2(0,d))) - int(input(p + i_int2(0,-d)))) +
::abs(int(input(p + i_int2(-d,0))) - int(input(p + i_int2(d,0))));
}
bc2s64_feature<A>::bc2s64_feature(const obox2d& d)
: bc2s_feature<A>(d)
{
for (unsigned i = 0; i < 16; i += 4)
{
point2d<int> p(10,10);
this->offsets_s1_[i/4] = (long(&this->s1_(p + i_int2(circle_r3[i]))) - long(&this->s1_(p))) / sizeof(V);
this->offsets_s2_[i/4] = (long(&this->s2_(p + i_int2(circle_r3[i])*2)) - long(&this->s2_(p))) / sizeof(V);
}
}
i_int2 host_image2d<V>::index_to_point(int idx) const
{
assert(!(pitch_ % sizeof(V)));
int r = (idx * int(sizeof(V))) / pitch_;
int c = idx - r * (pitch_ / int(sizeof(V)));
return i_int2(r, c);
}
inline __host__ __device__
i_short2 gradient_descent_match2(i_short2 prediction, F f, FI& feature_img, float& distance, unsigned scale = 1)
{
i_short2 match = prediction;
float match_distance = feature_img.distance(f, prediction);
unsigned match_i = 8;
box2d domain = feature_img.domain() - border(7);
if (!domain.has(prediction)) return prediction;
assert(domain.has(prediction));
for (int search = 0; search < 7; search++)
{
for(int i = 0; i != 25; i++)
{
i_int2 n(prediction + i_int2(c25_h[i]));
{
float d = feature_img.distance(f, n, scale);
if (d < match_distance)
{
match = n;
match_i = i;
match_distance = d;
}
}
}
if (i_int2(prediction) == i_int2(match) || !domain.has(match))
break;
else
prediction = match;
}
distance = match_distance;
return match;
}
__global__ void convolve_kernel(I, kernel_image2d<O> out, unsigned kernelsize)
{
i_int2 p = thread_pos2d();
if (!out.has(p))
return;
bt_change_vtype(O, type_mult(bt_vtype(O), float)) r = zero();
for(int i = 0; i < kernelsize; i++)
{
float w = tex1Dfetch(tex_weights, i);
point2d<int> n = i_int2(tex1Dfetch(tex_dpoints, i)) + p;
if (out.has(n))
r += O(tex2D(conv_input_tex<I>::tex(), n)) * w;
}
out(p) = r;
}
inline __host__ __device__
i_int2 naive_local_match(i_short2 prediction,
const S& sample,
const F& feature_img)
{
i_short2 match = prediction;
float match_distance = feature_img.distance(sample, prediction);
for_all_in_static_neighb2d(prediction, n, c49)
if (n->row() > 10 && n->row() < (feature_img.domain().nrows() - 10) &&
n->col() > 10 && n->col() < (feature_img.domain().ncols() - 10))
{
float d = feature_img.distance_linear(sample, n);
if (d < match_distance)
{
match = i_int2(n);
match_distance = d;
}
}
return match;
}
inline __host__ __device__
gl8u compute_saliency(i_short2 p, I& input_, int n_, int fast_th_, const A&)
{
unsigned max_n = 0;
unsigned n_bright = 0;
unsigned n_dark = 0;
gl8u vp = input_(p);
unsigned sum_bright = 0;
unsigned sum_dark = 0;
for (int i = 0; i < 16; i++)
{
gl8u vn = input_(p + i_int2(arch_neighb2d<A>::get(circle_r3_h, circle_r3, i)));
unsigned char dist = ::abs(int(vn) - int(vp));
if (dist > fast_th_)
{
if (int(vn) >= int(vp))
{
n_bright++;
sum_bright += dist;
}
else
{
n_dark++;
sum_dark += dist;
}
}
}
max_n = std::max(n_dark, n_bright);
if (max_n < n_)
return 0;
else
return max(sum_bright, sum_dark);
}
__host__ __device__ void benchmark_matcher_kernel(thread_info<target> ti,
i_short2* particles_vec,
unsigned n_particles,
F f,
kernel_image2d<T> particles,
kernel_image2d<T> new_particles,
kernel_image2d<i_short2> matches,
kernel_image2d<typename F::feature_t> states,
const kernel_image2d<i_short2> ls_matches,
i_short2 mvt,
T* compact_particles
,kernel_image2d<i_float4> dist
)
{
unsigned threadid = ti.blockIdx.x * ti.blockDim.x + ti.threadIdx.x;
if (threadid >= n_particles) return;
point2d<int> p = particles_vec[threadid];
if (!particles.has(p))
return;
matches(p) = i_short2(-1, -1);
if (p.row() < 6 || p.row() >= particles.domain().nrows() - 6 ||
p.col() < 6 || p.col() >= particles.domain().ncols() - 6)
{
new_particles(p).age = 0;
return;
}
if (particles(p).age == 0)
return;
// Search for the best match in the neighborhood.
short p_age = particles(p).age;
float p_pertinence = particles(p).pertinence;
typename F::feature_t p_state = states(p);
point2d<int> match = p;
bool ls_p_found = false;
float match_distance;
point2d<int> prediction = i_int2(p);
if (p_age == 1)
{
point2d<int> lsp(p.row() / 2, p.col() / 2);
point2d<int> ls_pred;
{for_all_in_static_neighb2d(lsp, n, c25)
{
if (ls_matches(n).x != -1)
{
ls_pred = point2d<int>(ls_matches(n).x * 2,
ls_matches(n).y * 2);
ls_p_found = true;
}
}
}
match = p;
match_distance = f.distance_linear(p_state, p);
}
void neighb_iterator2d<N>::start()
{
i_ = 0;
p_ = i_int2(cur_) + i_int2(n_[i_]);
i_++;
}
void neighb_iterator2d<N>::next()
{
p_ = i_int2(cur_) + i_int2(n_[i_]);
i_++;
}
inline __host__ __device__
gl8u compute_saliency(i_short2 p, I& input_, int n_, int fast_th_, const A&)
{
unsigned max_n = 0;
unsigned n = 0;
int status = 2;
gl8u vp = input_(p);
unsigned sum_bright = 0;
unsigned sum_dark = 0;
// if (n_ == 15)
// {
// int to_test[3] = {0, 4, 12, 0};
// for (int i = 0; i < 3; i++)
// {
// gl8u vn = input_(p + i_int2(arch_neighb2d<A>::get(circle_r3_h, circle_r3, to_test[i])));
// int sign = int(vn) > int(vp);
// unsigned char dist = ::abs(int(vn) - int(vp));
// if (dist > fast_th_)
// if (sign == status) n++;
// else
// {
// if (n > max_n)
// max_n = n;
// status = sign;
// n = 1;
// }
// else
// {
// if (n > max_n)
// max_n = n;
// status = 2;
// }
// }
// if (n < 3) return 0;
// }
if (n_ >= 9)
{
int to_test[6] = {0, 4, 8, 12, 0, 4};
int equals = 0;
for (int i = 0; i < 6; i++)
{
gl8u vn = input_(p + i_int2(arch_neighb2d<A>::get(circle_r3_h, circle_r3, to_test[i])));
int sign = int(vn) > int(vp);
unsigned char dist = ::abs(int(vn) - int(vp));
if (dist > fast_th_)
if (sign == status) { n++; }
else
{
if (n > max_n)
max_n = n;
status = sign;
n = 1;
}
else
{
if (n > max_n)
max_n = n;
status = 2;
equals++;
if (i < 4 && n_ >= 12 && equals >= 2) return 0;
}
}
if (n > max_n)
max_n = n;
if (n_ >= 12 && max_n < 3) return 0;
else if (n_ >= 8 && max_n < 2) return 0;
}
n = 0;
status = 2;
max_n = 0;
for (int i = 0; i < 16; i++)
{
gl8u vn = input_(p + i_int2(arch_neighb2d<A>::get(circle_r3_h, circle_r3, i)));
int sign = int(vn) > int(vp);
unsigned char dist = ::abs(int(vn) - int(vp));
if (dist > fast_th_)
{
if (sign == status) n++;
else
{
if (n > max_n)
max_n = n;
status = sign;
n = 1;
}
if (vn < vp)
sum_dark += dist;
else
sum_bright += dist;
}
else
{
if (n > max_n)
max_n = n;
status = 2;
}
}
if (n != 16 && status != 2)
{
int i = 0;
while (true)
{
gl8u vn = input_(p + i_int2(arch_neighb2d<A>::get(circle_r3_h, circle_r3, i)));
int sign = int(vn) > int(vp);
unsigned char dist = ::abs(int(vn) - int(vp));
if (dist <= fast_th_ || sign != status) break;
n++;
i++;
assert(i < 16);
}
}
if (n > max_n)
max_n = n;
if (max_n < n_)
return 0;
else
return max(sum_bright, sum_dark);
// return sum_bright;
}
inline __host__ __device__
std::pair<i_short2, float> gradient_descent_match(i_short2 prediction_, F f, FI& feature_img, unsigned scale = 1)
{
typedef std::pair<i_short2, float> ret;
i_short2 prediction = prediction_;
typedef typename FI::architecture A;
i_short2 match = prediction;
float match_distance = feature_img.distance(f, prediction);
unsigned match_i = 8;
box2d domain = feature_img.domain() - border(0);
if (!domain.has(prediction))
{
return ret(prediction, 999999.f);
}
assert(domain.has(prediction));
for (int search = 0; search < 10; search++)
{
int i = arch_neighb2d<A>::get(c8_it_h, c8_it, match_i)[0];
int end = arch_neighb2d<A>::get(c8_it_h, c8_it, match_i)[1];
{
i_int2 n(prediction + i_int2(arch_neighb2d<A>::get(c8_h, c8, i)));
if (feature_img.domain().has(n))
{
float d = feature_img.distance(f, n, scale);
if (d < match_distance)
{
match = n;
match_i = i;
match_distance = d;
}
}
i = (i + 1) & 7;
}
#pragma unroll 4
for(; i != end; i = (i + 1) & 7)
{
i_int2 n(prediction + i_int2(arch_neighb2d<A>::get(c8_h, c8, i)));
if (feature_img.domain().has(n))
{
float d = feature_img.distance(f, n, scale);
if (d < match_distance)
{
match = n;
match_i = i;
match_distance = d;
}
}
}
if (i_int2(prediction) == i_int2(match) || !domain.has(match))
break;
else
prediction = match;
}
// if (scale == 2)
// match = prediction + 2 * (match - prediction);
// match_distance = feature_img.distance(f, match, scale);
return ret(match, match_distance);
}