void compute_bow_histogram(cv::Mat &sample, Histogram &feature_vector, cv::Ptr<cv::FeatureDetector> &detector, cv::SiftDescriptorExtractor &extractor, Quantization *quant){
//detect keypoints
std::vector<cv::KeyPoint> keypoints;
detector->detect( sample, keypoints );
//compute descriptor
cv::Mat descriptor_uchar;
extractor.compute(sample, keypoints, descriptor_uchar);
cv::Mat descriptor_double;
descriptor_uchar.convertTo(descriptor_double, CV_64F);
//convert from mat to bag of unquantized features
BagOfFeatures unquantized_features;
convert_mat_to_vector(descriptor_double, unquantized_features);
//quantize regions -- true BagOfFeatures
quant->quantize(unquantized_features, feature_vector);
}
Histogram compute_histogram(std::string imFile, cv::Ptr<cv::FeatureDetector> &detector, cv::Ptr<cv::FeatureDetector> &detector2, cv::SiftDescriptorExtractor &extractor, Quantization *quant) {
cv::Mat img = cv::imread(imFile);
//detect SIFT keypoints
std::vector<cv::KeyPoint> keypoints;
detector->detect( img, keypoints );
//detect MSER keypoints
std::vector<cv::KeyPoint> keypoints2;
detector2->detect( img, keypoints2 );
//group them together
for(cv::KeyPoint& keypoint : keypoints2) {
keypoints.push_back(keypoint);
}
std::cout << " - keypoint_ct: " << keypoints.size() << std::endl;
//compute descriptors
cv::Mat descriptor_uchar;
extractor.compute(img, keypoints, descriptor_uchar);
cv::Mat descriptor_double;
descriptor_uchar.convertTo(descriptor_double, CV_64F);
//convert from mat to bag of unquantized features
BagOfFeatures unquantized_features;
convert_mat_to_vector(descriptor_double, unquantized_features);
//quantize to form bag of words
Histogram bag_of_words;
quant->quantize(unquantized_features, bag_of_words);
//normalize
return bag_of_words;
}
//gets centroid for category from training images
void LocalDescriptorAndBagOfFeature::train_category(const std::vector<cv::Mat> &samples, Histogram ¢roid, const cv::Ptr<cv::FeatureDetector> &detector, const cv::SiftDescriptorExtractor &extractor, Quantization *quant){
clock_t start = clock();
int i = 0;
for(const cv::Mat& sample : samples){
i++;
std::cout << "converting img " << i << " of " << samples.size() << " to bag of features" << std::endl;
//detect keypoints
std::vector<cv::KeyPoint> keypoints;
detector->detect( sample, keypoints );
//compute descriptor
cv::Mat descriptor_uchar;
extractor.compute(sample, keypoints, descriptor_uchar);
cv::Mat descriptor_double;
descriptor_uchar.convertTo(descriptor_double, CV_64F);
//convert from mat to bag of unquantized features
BagOfFeatures unquantized_features;
convert_mat_to_vector(descriptor_double, unquantized_features);
//quantize regions -- true BagOfFeatures
Histogram feature_vector;
quant->quantize(unquantized_features, feature_vector);
//aggregate
vector_add(centroid, feature_vector);
}
//divide by training category size to compute centroid
//std::transform(centroid.begin(), centroid.end(), centroid.begin(), std::bind1st(std::divides<double>(),bikes.size()));
for(double& d : centroid){
d = d/samples.size();
}
std::cout << double( clock() - start ) / (double)CLOCKS_PER_SEC<< " seconds." << std::endl;
}
/* DoGカーネルによる特徴量の抽出 */
void SIFTExtractor::extract_using_dense()
{
/*+ パラメータの算出 ++++++++++++++++*/
Size image_size = Size(proc_image.cols, proc_image.rows); // 画像サイズの取得
// 1. サンプリング間隔を求める
double interval = sqrt((image_size.width * image_size.height) / (double)feature_num);
// 2. スケールの決定（サンプリング間隔 interval / 2.0)
double scale = interval / 2.0;
// 3. interval間隔で横・縦ともに抽出できるサンプル数の算出
int sample_col_num = (double)image_size.width / floor(interval);
int sample_row_num = (double)image_size.height / floor(interval);
// 4. 画像の余白を算出
int odd_cols, odd_rows;
if( (image_size.width % sample_col_num == 0) && (image_size.height % sample_row_num == 0) )
{
//cout << "1" << endl;
odd_cols = image_size.width - ( (sample_col_num - 1) * floor(interval) );
odd_rows = image_size.height - ( (sample_row_num - 1) * floor(interval) );
}
else if( (image_size.width % sample_col_num == 0) && (image_size.height % sample_row_num != 0) )
{
//cout << "2" << endl;
odd_cols = image_size.width - ( (sample_col_num - 1) * floor(interval) );
odd_rows = image_size.height - ( sample_row_num * floor(interval) );
}
else if( (image_size.width % sample_col_num == 0) && (image_size.height % sample_row_num == 0) )
{
//cout << "3" << endl;
odd_cols = image_size.width - ( sample_col_num * floor(interval) );
odd_rows = image_size.height - ( (sample_row_num - 1) * floor(interval) );
}
else
{
//cout << "4" << endl;
odd_cols = image_size.width - ( sample_col_num * floor(interval) );
odd_rows = image_size.height - ( sample_row_num * floor(interval) );
}
// 5. 左上から何pixelシフトした位置から特徴量を抽出するかを算出
double sift = sqrt((odd_cols * odd_rows) / 4.0);
// 6. 正方形の画像の場合の例外処理
if( image_size.width == image_size.height)
{
int tmp = (double)image_size.width - (floor(sift) * 2.0) - (floor(interval) * (sqrt(feature_num) - 1.0));
if(tmp == 0)
sift--;
else
sift += tmp/2;
}
/*
cout << "Size (" << proc_image.cols << " * " << proc_image.rows << ")" << endl;
cout << "interval:" << interval << endl;
cout << "scale:" << scale << endl;
cout << "sample_col_num:" << sample_col_num << endl;
cout << "sample_row_num:" << sample_row_num << endl;
cout << "odd_cols:" << odd_cols << endl;
cout << "odd_rows:" << odd_rows << endl;
cout << "sift:" << sift << endl;
*/
/*+ パラメータの算出 ++++++++++++++++*/
// 1. キーポイントの抽出
while(true)
{
cv::DenseFeatureDetector detector(floor(scale), 1, 0.1f, floor(interval), floor(sift), true, false);
detector.detect(proc_image, keypoints);
// 指定した特徴量になるまでフィードバックループ
if(keypoints.size() < feature_num)
{
sift--;
if(sift < 0)
cerr << "計算式がおかしい" << endl;
}
else if(keypoints.size() > feature_num)
{
sift++;
if(sift < 0)
cerr << "計算式がおかしい" << endl;
}
else
break;
}
// 2. SIFT記述子の抽出
extractor.compute(proc_image, keypoints, descriptors);
//TODO: エラー処理
// 抽出済みフラグをセット
extraction = true;
}