bool saveToDirectory(string directory, int i_update, const DistributionGaussian& distribution, const Rollout* rollout_eval, const vector<Rollout*>& rollouts, const VectorXd& weights, const DistributionGaussian& distribution_new, bool overwrite)
{
vector<DistributionGaussian> distribution_vec;
distribution_vec.push_back(distribution);
vector<DistributionGaussian> distribution_new_vec;
distribution_new_vec.push_back(distribution_new);
return saveToDirectory(directory, i_update, distribution_vec, rollout_eval, rollouts, weights, distribution_new_vec, overwrite);
}
void runEvolutionaryOptimization(
const CostFunction* const cost_function,
const DistributionGaussian* const initial_distribution,
const Updater* const updater,
int n_updates,
int n_samples_per_update,
string save_directory,
bool overwrite,
bool only_learning_curve)
{
// Some variables
MatrixXd samples;
VectorXd costs, cost_eval;
// Bookkeeping
vector<UpdateSummary> update_summaries;
UpdateSummary update_summary;
if (save_directory.empty())
cout << "init = " << " distribution=" << *initial_distribution;
// Optimization loop
DistributionGaussian* distribution = initial_distribution->clone();
for (int i_update=0; i_update<n_updates; i_update++)
{
// 0. Get cost of current distribution mean
cost_function->evaluate(distribution->mean().transpose(),cost_eval);
// 1. Sample from distribution
distribution->generateSamples(n_samples_per_update, samples);
// 2. Evaluate the samples
cost_function->evaluate(samples,costs);
// 3. Update parameters
updater->updateDistribution(*distribution, samples, costs, *distribution, update_summary);
// Some output and/or saving to file (if "directory" is set)
if (save_directory.empty())
{
cout << "\t cost_eval=" << cost_eval << endl << i_update+1 << " " << *distribution;
}
else
{
update_summary.cost_eval = cost_eval[0];
update_summaries.push_back(update_summary);
}
}
// Save update summaries to file, if necessary
if (!save_directory.empty())
saveToDirectory(update_summaries,save_directory,overwrite,only_learning_curve);
}
void Package::saveAs(const String& name, bool remove_unused) {
// type of package
if (wxDirExists(name)) {
saveToDirectory(name, remove_unused, false);
} else {
saveToZipfile (name, remove_unused, false);
}
filename = name;
removeTempFiles(remove_unused);
reopen();
}
bool saveToDirectoryNewUpdate(const UpdateSummary& update_summary, std::string directory)
{
// Make directory if it doesn't already exist
if (!boost::filesystem::exists(directory))
{
if (!boost::filesystem::create_directories(directory))
{
cerr << __FILE__ << ":" << __LINE__ << ":";
cerr << "Couldn't make directory file '" << directory << "'." << endl;
return false;
}
// Directory didn't exist yet, so this must be the first update.
directory += "/update00001/";
bool overwrite=true;
return saveToDirectory(update_summary,directory,overwrite);
}
// Find the directory with the highest update
// todo: this can probably be done more efficiently with boost::filesystem somehow
int MAX_UPDATE = 99999; // Cannot store more in %05d format
int i_update=1;
while (i_update<MAX_UPDATE)
{
stringstream stream;
stream << directory << "/update" << setw(5) << setfill('0') << i_update << "/";
string directory_update = stream.str();
if (!boost::filesystem::exists(directory_update))
{
// Found a directory that doesn't exist yet!
bool overwrite=true;
return saveToDirectory(update_summary,directory_update,overwrite);
}
i_update++;
}
std::cerr << "Sorry, directory " << directory << " is already full with update subdirectories." << std::endl;
return false;
}
bool saveToDirectory(const vector<UpdateSummaryParallel>& update_summaries, std::string directory, bool overwrite, bool only_learning_curve)
{
// Save the learning curve
int n_updates = update_summaries.size();
assert(n_updates>0);
int n_parallel = update_summaries[0].distributions.size();
MatrixXd learning_curve(n_updates,2+n_parallel);
learning_curve(0,0) = 0; // First evaluation is at 0
for (int i_update=0; i_update<n_updates; i_update++)
{
// Number of samples at which an evaluation was performed.
if (i_update>0)
{
int n_samples = update_summaries[i_update].costs.rows();
learning_curve(i_update,0) = learning_curve(i_update-1,0) + n_samples;
}
// The cost of the evaluation at this update
learning_curve(i_update,1) = update_summaries[i_update].cost_eval;
for (int i_parallel=0; i_parallel<n_parallel; i_parallel++)
{
// The largest eigenvalue of the covariance matrix, for each distribution
DistributionGaussian* distribution = update_summaries[i_update].distributions[i_parallel];
MatrixXd eigen_values = distribution->covar().eigenvalues().real();
learning_curve(i_update,2+i_parallel) = sqrt(eigen_values.maxCoeff());
}
}
if (!saveMatrix(directory, "learning_curve.txt", learning_curve, overwrite))
return false;
if (!only_learning_curve)
{
// Save all the information in the update summaries
for (int i_update=0; i_update<n_updates; i_update++)
{
stringstream stream;
stream << directory << "/update" << setw(5) << setfill('0') << i_update+1 << "/";
if (!saveToDirectory(update_summaries[i_update], stream.str(),overwrite))
return false;
}
}
return true;
}
bool saveToDirectory(string directory, int i_update, const vector<DistributionGaussian>& distribution, const Rollout* rollout_eval, const vector<Rollout*>& rollouts, const VectorXd& weights, const vector<DistributionGaussian>& distribution_new, bool overwrite)
{
VectorXd cost_eval;
if (rollout_eval!=NULL)
rollout_eval->cost(cost_eval);
MatrixXd costs(rollouts.size(),rollouts[0]->getNumberOfCostComponents());
for (unsigned int ii=0; ii<rollouts.size(); ii++)
{
VectorXd cur_cost;
rollouts[ii]->cost(cur_cost);
costs.row(ii) = cur_cost;
}
// Save update information
MatrixXd samples;
saveToDirectory(directory, i_update, distribution, cost_eval, samples, costs, weights, distribution_new,overwrite);
stringstream stream;
stream << directory << "/update" << setw(5) << setfill('0') << i_update << "/";
string directory_update = stream.str();
// Save rollouts too
for (unsigned int i_rollout=0; i_rollout<rollouts.size(); i_rollout++)
{
stringstream stream;
stream << directory_update << "/rollout" << setw(3) << setfill('0') << i_rollout+1;
if (!rollouts[i_rollout]->saveToDirectory(stream.str(),overwrite))
return false;
}
if (rollout_eval!=NULL)
if (rollout_eval->saveToDirectory(directory_update+"/rollout_eval",overwrite))
return false;
return true;
}
// This function could have been integrated with the above. But I preferred to duplicate a bit of
// code so that the difference between running an optimziation with a CostFunction or
// Task/TaskSolver is more apparent.
void runEvolutionaryOptimization(
const Task* const task,
const TaskSolver* const task_solver,
const DistributionGaussian* const initial_distribution,
const Updater* const updater,
int n_updates,
int n_samples_per_update,
string save_directory,
bool overwrite,
bool only_learning_curve)
{
// Some variables
MatrixXd samples;
MatrixXd cost_vars, cost_vars_eval;
VectorXd costs, cost_eval;
// Bookkeeping
vector<UpdateSummary> update_summaries;
UpdateSummary update_summary;
if (save_directory.empty())
cout << "init = " << " distribution=" << *initial_distribution;
// Optimization loop
DistributionGaussian* distribution = initial_distribution->clone();
for (int i_update=0; i_update<n_updates; i_update++)
{
// 0. Get cost of current distribution mean
task_solver->performRollouts(distribution->mean().transpose(),cost_vars_eval);
task->evaluate(cost_vars_eval,cost_eval);
// 1. Sample from distribution
distribution->generateSamples(n_samples_per_update, samples);
// 2A. Perform the roll-outs
task_solver->performRollouts(samples,cost_vars);
// 2B. Evaluate the samples
task->evaluate(cost_vars,costs);
// 3. Update parameters
updater->updateDistribution(*distribution, samples, costs, *distribution, update_summary);
// Some output and/or saving to file (if "directory" is set)
if (save_directory.empty())
{
cout << "\t cost_eval=" << cost_eval << endl << i_update+1 << " " << *distribution;
}
else
{
update_summary.cost_eval = cost_eval[0];
update_summary.cost_vars_eval = cost_vars_eval;
update_summary.cost_vars = cost_vars;
update_summaries.push_back(update_summary);
}
}
// Save update summaries to file, if necessary
if (!save_directory.empty())
{
saveToDirectory(update_summaries,save_directory,overwrite,only_learning_curve);
// If you store only the learning curve, no need to save the script to visualize rollouts
if (!only_learning_curve)
task->savePerformRolloutsPlotScript(save_directory);
}
}
/** \todo Get rid of runOptimizationParallelDeprecated(), and implement in UpdaterCovarAdapation
*/
void runOptimizationParallelDeprecated(
Task* task,
TaskSolver* task_solver,
vector<DistributionGaussian*> initial_distributions,
Updater* updater,
int n_updates,
int n_samples_per_update,
string save_directory,
bool overwrite,
bool only_learning_curve)
{
// Some variables
int n_parallel = initial_distributions.size();
assert(n_parallel>=2);
int n_samples = n_samples_per_update; // Shorthand
VectorXi offsets(n_parallel+1);
offsets[0] = 0;
for (int ii=0; ii<n_parallel; ii++)
offsets[ii+1] = offsets[ii] + initial_distributions[ii]->mean().size();
int sum_n_dims = offsets[n_parallel];
// n_parallel X n_samples X n_dims
// Note: n_samples must be the same for all, n_dims varies
//vector<MatrixXd> sample(n_parallel);
//for (int ii=0; ii<n_parallel; ii++)
// // Pre-allocate memory just to be clear.
// sample[ii] = MatrixXd(n_samples_per_update,initial_distributions[ii]->mean().size());
MatrixXd samples(n_samples,sum_n_dims);
// Some variables
VectorXd sample_eval(sum_n_dims);
VectorXd cost_eval;
MatrixXd cost_vars_eval;
MatrixXd samples_per_parallel;
MatrixXd cost_vars;
VectorXd cur_costs;
VectorXd costs(n_samples);
VectorXd total_costs(n_samples);
VectorXd weights;
// Bookkeeping
MatrixXd learning_curve(n_updates,3);
vector<DistributionGaussian> distributions;
vector<DistributionGaussian> distributions_new;
for (int ii=0; ii<n_parallel; ii++)
{
distributions.push_back(*(initial_distributions[ii]->clone()));
distributions_new.push_back(*(initial_distributions[ii]->clone()));
}
// Optimization loop
for (int i_update=0; i_update<n_updates; i_update++)
{
// 0. Get cost of current distribution mean
for (int pp=0; pp<n_parallel; pp++)
sample_eval.segment(offsets[pp],offsets[pp+1]-offsets[pp]) = distributions[pp].mean().transpose();
task_solver->performRollout(sample_eval,cost_vars_eval);
task->evaluateRollout(cost_vars_eval,sample_eval,cost_eval);
Rollout* rollout_eval = new Rollout(sample_eval,cost_vars_eval,cost_eval);
// 1. Sample from distribution
for (int pp=0; pp<n_parallel; pp++)
{
distributions[pp].generateSamples(n_samples, samples_per_parallel);
int width = offsets[pp+1]-offsets[pp];
samples.block(0,offsets[pp],n_samples,width) = samples_per_parallel;
}
vector<Rollout*> rollouts(n_samples_per_update);
for (int i_sample=0; i_sample<n_samples_per_update; i_sample++)
{
// 2. Perform rollouts for the samples
task_solver->performRollout(samples.row(i_sample), cost_vars);
// 3. Evaluate the last batch of rollouts
task->evaluateRollout(cost_vars,samples.row(i_sample),cur_costs);
// Bookkeeping
costs[i_sample] = cur_costs[0];
rollouts[i_sample] = new Rollout(samples.row(i_sample),cost_vars,cur_costs);
}
// 4. Update parameters
for (int pp=0; pp<n_parallel; pp++)
{
int width = offsets[pp+1]-offsets[pp];
samples_per_parallel = samples.block(0,offsets[pp],n_samples,width);
updater->updateDistribution(distributions[pp], samples_per_parallel, costs, weights, distributions_new[pp]);
}
// Some output and/or saving to file (if "directory" is set)
if (save_directory.empty())
{
cout << i_update+1 << " cost_eval=" << cost_eval << endl;
}
else
{
// Update learning curve
// How many samples so far?
learning_curve(i_update,0) = i_update*n_samples_per_update;
// Cost of evaluation
learning_curve(i_update,1) = cost_eval[0];
// Exploration magnitude
learning_curve(i_update,2) = 0.0;
for (int pp=0; pp<n_parallel; pp++)
learning_curve(i_update,2) += sqrt(distributions[pp].maxEigenValue());
// Save more than just learning curve.
if (!only_learning_curve)
{
saveToDirectory(save_directory,i_update,distributions,rollout_eval,rollouts,weights,distributions_new);
if (i_update==0)
task->savePlotRolloutScript(save_directory);
}
}
// Distribution is new distribution
for (int ii=0; ii<n_parallel; ii++)
distributions[ii] = distributions_new[ii];
}
}
void runOptimizationTask(
const Task* const task,
const TaskSolver* const task_solver,
const DistributionGaussian* const initial_distribution,
const Updater* const updater,
int n_updates,
int n_samples_per_update,
std::string save_directory,
bool overwrite,
bool only_learning_curve)
{
int n_cost_components = task->getNumberOfCostComponents();
// Some variables
VectorXd sample_eval;
MatrixXd cost_vars_eval;
VectorXd cost_eval(1+n_cost_components);
MatrixXd samples;
MatrixXd cost_vars;
VectorXd weights;
MatrixXd costs(n_samples_per_update,1+n_cost_components);
// tmp variables
VectorXd total_costs(n_samples_per_update);
VectorXd cur_cost(1+n_cost_components);
// Bookkeeping
MatrixXd learning_curve(n_updates,2+n_cost_components);
MatrixXd exploration_curve(n_updates,2);
if (save_directory.empty())
cout << "init = " << " distribution=" << *initial_distribution;
DistributionGaussian distribution = *(initial_distribution->clone());
DistributionGaussian distribution_new = *(initial_distribution->clone());
// Optimization loop
for (int i_update=0; i_update<n_updates; i_update++)
{
// 0. Get cost of current distribution mean
sample_eval = distribution.mean().transpose();
task_solver->performRollout(sample_eval,cost_vars_eval);
task->evaluateRollout(cost_vars_eval,sample_eval,cost_eval);
Rollout* rollout_eval = new Rollout(sample_eval,cost_vars_eval,cost_eval);
// 1. Sample from distribution
distribution.generateSamples(n_samples_per_update, samples);
vector<Rollout*> rollouts(n_samples_per_update);
for (int i_sample=0; i_sample<n_samples_per_update; i_sample++)
{
// 2A. Perform the rollout
task_solver->performRollout(samples.row(i_sample),cost_vars);
// 2B. Evaluate the rollout
task->evaluateRollout(cost_vars,samples.row(i_sample),cur_cost);
costs.row(i_sample) = cur_cost;
rollouts[i_sample] = new Rollout(samples.row(i_sample),cost_vars,cur_cost);
}
// 3. Update parameters (first column of costs contains sum of cost components)
total_costs = costs.col(0);
updater->updateDistribution(distribution, samples, total_costs, weights, distribution_new);
// Bookkeeping
// Some output and/or saving to file (if "directory" is set)
if (save_directory.empty())
{
cout << "\t cost_eval=" << cost_eval << endl << i_update+1 << " " << distribution;
}
else
{
// Update learning curve
// How many samples?
int i_samples = i_update*n_samples_per_update;
learning_curve(i_update,0) = i_samples;
// Cost of evaluation
learning_curve.block(i_update,1,1,1+n_cost_components) = cost_eval.transpose();
// Exploration magnitude
exploration_curve(i_update,0) = i_samples;
exploration_curve(i_update,1) = sqrt(distribution.maxEigenValue());
// Save more than just learning curve.
if (!only_learning_curve)
{
saveToDirectory(save_directory,i_update,distribution,rollout_eval,rollouts,weights,distribution_new);
if (i_update==0)
task->savePlotRolloutScript(save_directory);
}
}
// Distribution is new distribution
distribution = distribution_new;
}
// Save learning curve to file, if necessary
if (!save_directory.empty())
{
// Todo: save cost labels also
saveMatrix(save_directory, "exploration_curve.txt",exploration_curve,overwrite);
saveMatrix(save_directory, "learning_curve.txt",learning_curve,overwrite);
}
}
int main(int argc, char *argv[]) {
SPConfig config = NULL; // hold configuration parameters
char config_filename[CONFIG_FILE_PATH_SIZE]; // the configuration file name
int knn; // the number of similar features in each image to find (spKNN from configuration file)
int num_of_similar_images_to_find; // the number of similar images (to the query) to find (from configuration file)
int split_method; // holds an int representing the split method: 0=RANDOM, 1= MAX_SPREAD, 2=INCREMENTAL
bool minGui = false; // value of the system variable MinimalGui
bool extraction_mode; // indicates if extraction mode on or off
int num_of_images = 0; // number of images in the directory given by the user in the configuration file
char** all_images_paths = NULL; // array with the paths to all the images
int last_extracted_feature = 0; // helper - holds the last feature extracted in order to free all extracted features on error
SPPoint** features_per_image = NULL; // helper - holds the features for each images
int* num_of_features_per_image = NULL; // holds number of features extracted for each image
char query_image[CONFIG_FILE_PATH_SIZE]; // the query image
SPPoint* query_features = NULL; // all query features
int query_num_of_features; // number of features in query image
KDTreeNode kd_tree = NULL; // array holds a KDTree for the images
int* closest_images = NULL; // array holds the spNumOfSimilarImages indexes of the closest images to the query image
int print_result; // holds the result of the call to PrintMinGuiFalse
int retval = 0; // return value - default 0 on success
char string_holder[CONFIG_FILE_PATH_SIZE]; // helper to hold strings
sp::ImageProc *improc = NULL;
SP_CONFIG_MSG msg;
int i;
int j;
int n;
// validate command line arguments:
// cmd line arguments are ok if there was no arguments specified (argc == 1) or two arguments specified ( -c and filname)
if (argc != 3 && argc != 1) {
printf(INVALID_CMD_LINE_MSG);
return -1;
}
if (argc == 1) {
strcpy(config_filename, DEFAULT_CONFIG_FILENAME);
config = spConfigCreate(config_filename, &msg);
if (msg == SP_CONFIG_CANNOT_OPEN_FILE) {
printf(ERROR_OPENING_DEFAULT_CONFIG_FILE_MSG, DEFAULT_CONFIG_FILENAME);
}
if (msg != SP_CONFIG_SUCCESS) {
retval = -1;
goto err; // error is printed inside spConfigCreate
}
}
else { // argc == 3
// check that second argument is the -c flag
if (strcmp(argv[1], CMD_LINE_CONFIG_FILENAME_FLAG) != 0) {
printf(INVALID_CMD_LINE_MSG);
retval = -1;
goto err;
}
strcpy(config_filename, argv[2]);
config = spConfigCreate(config_filename, &msg);
if (msg == SP_CONFIG_CANNOT_OPEN_FILE) {
printf(ERROR_OPENING_CONFIG_FILE_MSG, config_filename);
}
if (msg != SP_CONFIG_SUCCESS) {
retval = -1;
goto err; // error is printed inside spConfigCreate
}
}
// initiate from config
if (initFromConfig(config, &num_of_images, &num_of_similar_images_to_find, &knn, &split_method, &extraction_mode, &minGui, &all_images_paths) == -1 ) {
retval = -1;
goto err; // error is printed inside initFromConfig
}
// initiate image proc
improc = new sp::ImageProc(config);
// extract images features
if ((num_of_features_per_image = (int*)malloc(sizeof(*num_of_features_per_image) * num_of_images)) == NULL) {
spLoggerPrintError(ALLOCATION_FAILURE_MSG, __FILE__, __func__, __LINE__);
retval = -1;
goto err;
}
spLoggerPrintInfo(CHECK_EXTRACTION_MODE_INFO_LOG);
if (extraction_mode) { // extraction mode is chosen
spLoggerPrintMsg(USE_EXTRACTION_MODE_LOG);
spLoggerPrintInfo(EXTRACT_IMAGES_FEATURES_INFO_LOG);
if ((features_per_image = (SPPoint**)malloc(sizeof(*features_per_image) * num_of_images)) == NULL) {
spLoggerPrintError(ALLOCATION_FAILURE_MSG, __FILE__, __func__, __LINE__);
retval = -1;
goto err;
}
// extract each image features and write them to file
for (i=0; i < num_of_images; i++) {
// extract image features
if ((features_per_image[i] = improc->getImageFeatures(all_images_paths[i], i, &(num_of_features_per_image[i]))) == NULL) {
last_extracted_feature = i;
retval = -1;
goto err; // error is printed inside getImageFeatures
}
}
if (saveToDirectory(config, features_per_image, num_of_features_per_image, num_of_images) == -1) {
retval = -1;
goto err; // error is printed inside saveToDirectory
}
}
else { // not extraction mode
spLoggerPrintMsg(USE_NOT_EXTRACTION_MODE_LOG);
spLoggerPrintInfo(READ_FEATURES_FROM_FILE_LOG);
if ((features_per_image = extractFromFiles(config, num_of_features_per_image, num_of_images)) == NULL) {
retval = -1;
goto err; // error is printed inside extractFromFiles
}
}
if ((kd_tree = initiateDataStructures(features_per_image, num_of_features_per_image, num_of_images, split_method)) == NULL) {
retval = -1;
goto err; // error is printed inside initiateDataStructures
}
query:
while(1) {
// get a query image from the user
printf(ENTER_AN_IMAGE_MSG);
fflush(NULL);
scanf("%s",query_image);
// exit if user asked
if (strcmp (query_image,EXIT_SIGN) == 0) {
printf(EXIT_MSG);
fflush(NULL);
goto err; // free memory and quit
}
if( access( query_image, F_OK ) == -1 ) {
printf(FILE_DOESNT_EXIST, query_image);
goto query;
}
// extract query image features
spLoggerPrintMsg(EXTRACT_QUERY_IMAGE_FEATURES_LOG);
if ((query_features = improc->getImageFeatures(query_image, num_of_images, &query_num_of_features)) == NULL) {
retval = -1;
goto err_inside_loop; // error log is printed inside getImageFeatures
}
// print debug log
if ((n = sprintf(string_holder, NUM_OF_EXTRACTED_FEATURES_DEBUG_LOG, query_num_of_features)) < 0) {
spLoggerPrintError(GENERAL_ERROR_MSG, __FILE__, __func__, __LINE__);
retval = -1;
goto err_inside_loop;
}
spLoggerPrintDebug(string_holder, __FILE__, __func__, __LINE__);
// print log message
if ((n = sprintf(string_holder, SEARCING_SIMILAR_IMAGES_MSG, num_of_similar_images_to_find)) < 0) {
spLoggerPrintError(GENERAL_ERROR_MSG, __FILE__, __func__, __LINE__);
retval = -1;
goto err_inside_loop;
}
spLoggerPrintMsg(string_holder);
// find similar images to the query image
closest_images = getKClosestImages(num_of_similar_images_to_find, knn, query_features,
kd_tree, query_num_of_features, num_of_images);
if (closest_images == NULL) {
retval = -1;
goto err_inside_loop; // error is printed to inside getKClosestImages
}
// show (display) closest_images images
//need to show images
if (minGui==true){
for (i=0; i<num_of_similar_images_to_find; i++){
//get file path of the images by the indexes in closest_images
improc->showImage(all_images_paths[closest_images[i]]);
}
}
// i.e. minGui==false, just need to print images path
else{
print_result = PrintMinGuiFalse(query_image, num_of_similar_images_to_find, all_images_paths, closest_images);
if (print_result == 0) {
retval = -1;
goto err_inside_loop; // error is printed inside
}
}
// free memory before entering the loop again
free(closest_images);
if (query_features != NULL) {
for (i=0; i<query_num_of_features; i++) {
spPointDestroy(query_features[i]);
}
free(query_features);
}
}
err_inside_loop:
free(closest_images);
// free query_features
if (query_features != NULL) {
for (i=0; i<query_num_of_features; i++) {
spPointDestroy(query_features[i]);
}
free(query_features);
}
// done - destroy logger and free everything
err:
spLoggerDestroy();
// free the kd tree
DestroyKDTreeNode(kd_tree);
spConfigDestroy(config);
// free all images paths
if (all_images_paths != NULL) {
for (i = 0; i < num_of_images; i ++) {
free(all_images_paths[i]);
}
free(all_images_paths);
}
if (features_per_image != NULL) {
// free features_per_image
for (i = 0; i < last_extracted_feature; i ++) {
if (features_per_image[i] != NULL) {
for (j = 0; j < num_of_features_per_image[i]; j++) {
spPointDestroy(features_per_image[i][j]);
}
}
}
free(features_per_image);
}
free(num_of_features_per_image); // must be freed after features_per_image
if (improc != NULL) {
delete improc;
}
return retval;
}
