/**
* Get the prediction response for the given image.
* @param originImage image, which should be predicted
* @param resultLayer the name of the result layer
* @param dataLayer the name of the data layer
* @param predictions the predictions
*/
void CaffeClassifier::predict(std::vector<cv::Mat> originImages, std::vector<int> labels, string resultLayer,
string dataLayer, vector<short> & predictions) {
vector<Datum> vecDatum;
for (int i = 0; i < originImages.size(); i++) {
cv::Mat originImage = originImages[i];
// resize image
Mat image;
if (originImage.cols != imageSize.width || originImage.rows != imageSize.height) {
resize(originImage, image, imageSize);
} else
image = originImage;
// check channels
if (channels != image.channels()) {
cerr << "Error: the channel number of input image is invalid for CNN classifier!" << endl;
exit(1);
}
// mat to datum
Datum datum;
CVMatToDatum(image, &datum);
datum.set_label(labels[i]);
vecDatum.push_back(datum);
image.release();
}
// get the data layer
const caffe::shared_ptr<MemoryDataLayer<float>> memDataLayer = boost::static_pointer_cast<MemoryDataLayer<float>> (caffeNet->layer_by_name(dataLayer));
// push new image data
memDataLayer->AddDatumVector(vecDatum);
//memDataLayer->ExactNumBottomBlobs();
// do forward pass
vector<Blob<float>*> inputVec;
caffeNet->Forward(inputVec);
// get results
const caffe::shared_ptr<Blob<float> > featureBlob = caffeNet->blob_by_name(resultLayer);
int batchSize = featureBlob->num();
int dimFeatures = featureBlob->count() / batchSize;
// std::cout << "Batch size is " << batchSize << "/ dim features is " << dimFeatures << std::endl;
// get output from each channel
for (int n = 0; n < batchSize; ++n) {
float* fs = featureBlob->mutable_cpu_data() + featureBlob->offset(n);
if (sizeof(fs) > 0) {
vector<float> feature_vector(fs, fs + dimFeatures);
predictions.insert(predictions.end(), feature_vector.begin(), feature_vector.end());
}
}
// release data
// for (Datum d : vecDatum) {
// d.release_data();
// }
}
void
extract_dense(nv_matrix_t *vlad, int j,
const nv_matrix_t *image,
nv_keypoint_dense_t *dense,
int ndense
)
{
NV_ASSERT(vlad->n == DIM);
int desc_m;
nv_matrix_t *key_vec;
nv_matrix_t *desc_vec;
nv_matrix_t *resize, *gray, *smooth;
int i;
int km = 0;
if (m_fit_area == 0) {
float scale = IMG_SIZE() / (float)NV_MAX(image->rows, image->cols);
resize = nv_matrix3d_alloc(3, (int)(image->rows * scale),
(int)(image->cols * scale));
} else {
float axis_ratio = (float)image->rows / image->cols;
int new_cols = (int)sqrtf(m_fit_area / axis_ratio);
int new_rows = (int)((float)m_fit_area / new_cols);
resize = nv_matrix3d_alloc(3, new_rows, new_cols);
}
gray = nv_matrix3d_alloc(1, resize->rows, resize->cols);
smooth = nv_matrix3d_alloc(1, resize->rows, resize->cols);
for (i = 0; i < ndense; ++i) {
km += dense[i].rows * dense[i].cols;
}
km *= 2;
key_vec = nv_matrix_alloc(NV_KEYPOINT_KEYPOINT_N, km);
desc_vec = nv_matrix_alloc(NV_KEYPOINT_DESC_N, km);
nv_resize(resize, image);
nv_gray(gray, resize);
nv_gaussian5x5(smooth, 0, gray, 0);
nv_matrix_zero(desc_vec);
nv_matrix_zero(key_vec);
desc_m = nv_keypoint_dense_ex(m_ctx, key_vec, desc_vec, smooth, 0,
dense, ndense);
feature_vector(vlad, j, key_vec, desc_vec, desc_m);
nv_matrix_free(&gray);
nv_matrix_free(&resize);
nv_matrix_free(&smooth);
nv_matrix_free(&key_vec);
nv_matrix_free(&desc_vec);
}
DenseMatrix dense_matrix_from_features(FeaturesCallback features,
IndexType dimension,
RandomAccessIterator begin,
RandomAccessIterator end)
{
DenseMatrix matrix(dimension, end-begin);
DenseVector feature_vector(dimension);
for (RandomAccessIterator iter=begin; iter!=end; ++iter)
{
features.vector(*iter,feature_vector);
matrix.col(iter-begin).array() = feature_vector;
}
return matrix;
}
// Helper computes one or more features corresponding to the given points.
// Emitted features are on the line defined by:
// start_pt + lambda * (end_pt - start_pt) for scalar lambda.
// Features are spaced at feature_length intervals.
static int ComputeFeatures(const FCOORD& start_pt, const FCOORD& end_pt,
double feature_length,
GenericVector<INT_FEATURE_STRUCT>* features) {
FCOORD feature_vector(end_pt - start_pt);
if (feature_vector.x() == 0.0f && feature_vector.y() == 0.0f) return 0;
// Compute theta for the feature based on its direction.
uint8_t theta = feature_vector.to_direction();
// Compute the number of features and lambda_step.
double target_length = feature_vector.length();
int num_features = IntCastRounded(target_length / feature_length);
if (num_features == 0) return 0;
// Divide the length evenly into num_features pieces.
double lambda_step = 1.0 / num_features;
double lambda = lambda_step / 2.0;
for (int f = 0; f < num_features; ++f, lambda += lambda_step) {
FCOORD feature_pt(start_pt);
feature_pt += feature_vector * lambda;
INT_FEATURE_STRUCT feature(feature_pt, theta);
features->push_back(feature);
}
return num_features;
}
