patch_kmeans2.diff
| modules/core/src/matrix.cpp (working copy) | ||
|---|---|---|
| 2196 | 2196 |
return d; |
| 2197 | 2197 |
} |
| 2198 | 2198 | |
| 2199 |
#ifdef HAVE_TBB |
|
| 2200 | ||
| 2201 |
class KMeansPPDistanceComputer |
|
| 2202 |
{
|
|
| 2203 |
float *tdist2; |
|
| 2204 |
const float *data; |
|
| 2205 |
const float *dist; |
|
| 2206 |
const int dims; |
|
| 2207 |
const int step; |
|
| 2208 |
const int stepci; |
|
| 2209 | ||
| 2210 |
public: |
|
| 2211 | ||
| 2212 |
KMeansPPDistanceComputer(float *_tdist2, |
|
| 2213 |
const float *_data, |
|
| 2214 |
const float *_dist, |
|
| 2215 |
int _dims, |
|
| 2216 |
int _step, |
|
| 2217 |
int _stepci) |
|
| 2218 |
: tdist2(_tdist2), |
|
| 2219 |
data(_data), |
|
| 2220 |
dist(_dist), |
|
| 2221 |
dims(_dims), |
|
| 2222 |
step(_step), |
|
| 2223 |
stepci(_stepci) |
|
| 2224 |
{
|
|
| 2225 | ||
| 2226 |
} |
|
| 2227 | ||
| 2228 |
void operator()(const cv::BlockedRange& range) const |
|
| 2229 |
{
|
|
| 2230 | ||
| 2231 |
const int begin = range.begin(); |
|
| 2232 |
const int end = range.end(); |
|
| 2233 | ||
| 2234 |
for (int i=begin; i<end; ++i) {
|
|
| 2235 |
tdist2[i] = std::min(distance(data + step*i, data + stepci, dims), dist[i]); |
|
| 2236 |
} |
|
| 2237 |
} |
|
| 2238 |
|
|
| 2239 |
}; |
|
| 2240 | ||
| 2241 |
#endif //HAVE_TBB |
|
| 2242 | ||
| 2243 | ||
| 2199 | 2244 |
/* |
| 2200 | 2245 |
k-means center initialization using the following algorithm: |
| 2201 | 2246 |
Arthur & Vassilvitskii (2007) k-means++: The Advantages of Careful Seeding |
| ... | ... | |
| 2227 | 2272 | |
| 2228 | 2273 |
for( j = 0; j < trials; j++ ) |
| 2229 | 2274 |
{
|
| 2230 |
double p = (double)rng*sum0, s = 0;
|
|
| 2275 |
double p = (double)rng*sum0; |
|
| 2231 | 2276 |
for( i = 0; i < N-1; i++ ) |
| 2232 | 2277 |
if( (p -= dist[i]) <= 0 ) |
| 2233 | 2278 |
break; |
| 2234 | 2279 |
int ci = i; |
| 2280 |
double s=0; |
|
| 2281 |
#ifndef HAVE_TBB |
|
| 2235 | 2282 |
for( i = 0; i < N; i++ ) |
| 2236 | 2283 |
{
|
| 2237 | 2284 |
tdist2[i] = std::min(distance(data + step*i, data + step*ci, dims), dist[i]); |
| 2238 | 2285 |
s += tdist2[i]; |
| 2239 | 2286 |
} |
| 2287 |
#else |
|
| 2288 |
cv::parallel_for(cv::BlockedRange(0, N), |
|
| 2289 |
KMeansPPDistanceComputer(tdist2, data, dist, dims, step, step*ci)); |
|
| 2290 |
|
|
| 2291 |
for( i = 0; i < N; i++ ) |
|
| 2292 |
{
|
|
| 2293 |
s += tdist2[i]; |
|
| 2294 |
} |
|
| 2295 |
#endif //HAVE_TBB |
|
| 2240 | 2296 |
|
| 2241 | 2297 |
if( s < bestSum ) |
| 2242 | 2298 |
{
|