Hello!
**Context:** I am currently attempting to train a SVM to recognize a specific building in a scene vs. different buildings. After attempting simpler approaches like direct image feature matching + homography and such, I want to attempt something a bit more flexible and/or powerful. Hence, I have set my gaze upon BOW+SVM.
**Problem:** As far as I am aware, the training phase goes without a hitch. However, when I reuse (either straight in code or after reloading the saved text format) my trained vocabulary and SVM to attempt a prediction on a new image, I get an out of memory error on the *svm.predict(descriptors)* function:
> *Unhandled exception at 0x7594C41F in MachineLearningTrainerBOW.exe:> Microsoft C++ exception: cv::Exception> at memory location 0x0037F288.*
I honestly don't know what's wrong, as I have followed examples setup in a similar manner to my current code and at a glance I do not seem to do anything outlandish or different than what they do:
https://gilscvblog.com/2013/08/23/bag-of-words-models-for-visual-categorization/
http://www.morethantechnical.com/2011/08/25/a-simple-object-classifier-with-bag-of-words-using-opencv-2-3-w-code/
http://answers.opencv.org/question/27138/setting-vocabulary-in-creating-bag-of-features-to-do-svm-classification/
http://answers.opencv.org/question/24650/how-to-2-class-categolization-using-surfbowsvm/
Could anyone point the way to a complete OpenCV tutorial or exemple covering the use of BOW+SVM, or maybe point out what I might be doing wrong from my code? Thank you!
// cluster count
const int CLUSTER_COUNT = 1000;
// create global bow vocabulary with TermCriteria as per tutorial
Ptr matcher = DescriptorMatcher::create("FlannBased");
Ptr extractor = SurfDescriptorExtractor::create();
BOWImgDescriptorExtractor bowDE(extractor, matcher);
Ptr detector = SURF::create(400);
BOWKMeansTrainer bowTrainer(CLUSTER_COUNT, TermCriteria(CV_TERMCRIT_EPS + CV_TERMCRIT_ITER, 10, FLT_EPSILON), 1, KMEANS_PP_CENTERS);
//
//... some irrelevant functions for loading files and such...
//
void ComputeBow(vector positives, vector negatives, string vocabularySave)
{
// get SURF descriptors and add to BOW each input files
cout << "Acquiring descriptors, this might take a while... ";
for (int i = 0; i < positives.size(); i++)
{
vector keypoints;
detector->detect(positives[i], keypoints);
Mat descriptors;
extractor->compute(positives[i], keypoints, descriptors);
if (!descriptors.empty()) bowTrainer.add(descriptors);
}
/*for (int i = 0; i < negatives.size(); i++)
{
vector keypoints;
detector->detect(negatives[i], keypoints);
Mat descriptors;
extractor->compute(negatives[i], keypoints, descriptors);
if (!descriptors.empty()) bowTrainer.add(descriptors);
}*/
cout << "Description complete!" << endl;
// Create the vocabulary with KMeans.
cout << "Clustering features, this might take a while... ";
Mat vocabulary;
vocabulary = bowTrainer.cluster();
bowDE.setVocabulary(vocabulary);
SaveVocabulary(vocabulary, vocabularySave);
cout << "Clustering complete!" << endl;
}
void TrainSVM(vector positiveMats, vector negativeMats, string svmSave, string vocabularySave)
{
//Setup the BOW
ComputeBow(positiveMats, negativeMats, vocabularySave);
// create training data positive and negative
Mat train, response;
cout << "Creating training sets, this might take a while... ";
for (int i = 0; i < positiveMats.size(); i++)
{
// set training data using BOWImgDescriptorExtractor
vector keypoints;
detector->detect(positiveMats[i], keypoints);
Mat descriptors;
bowDE.compute(positiveMats[i], keypoints, descriptors);
if (!descriptors.empty())
{
train.push_back(descriptors); // update training data
response.push_back(1); // update response data
}
}
for (int i = 0; i < negativeMats.size(); i++)
{
// set training data using BOWImgDescriptorExtractor
vector keypoints;
detector->detect(negativeMats[i], keypoints);
Mat descriptors;
bowDE.compute(negativeMats[i], keypoints, descriptors);
if (!descriptors.empty())
{
train.push_back(descriptors); // update training data
response.push_back(-1); // update response data
}
}
cout << "Training... ";
// setup svm as per tutorial values
Ptr svm = SVM::create();
svm->setType(ml::SVM::C_SVC);
svm->setKernel(ml::SVM::RBF);
svm->setGamma(8);
svm->setDegree(10);
svm->setCoef0(1);
svm->setC(10);
svm->setNu(0.5);
svm->setP(0.1);
svm->setTermCriteria(cvTermCriteria(CV_TERMCRIT_EPS, 1000, FLT_EPSILON));
svm->train(train, ROW_SAMPLE, response);
svm->save(svmSave);
cout << "Training complete!" << endl;
cout << "SVM saved at " << svmSave << endl;
cout << "Vocabulary matrix saved at " << vocabularySave << endl;
}
int Predict(string queryLoad, string svmLoad, string vocabularyLoad)
{
// load image
Mat img = cv::imread(queryLoad, IMREAD_GRAYSCALE);
// load vocabulary data
Mat vocabulary = LoadVocabulary(vocabularyLoad);
// Set the vocabulary
bowDE.setVocabulary(vocabulary);
vector keypoints;
detector->detect(img, keypoints);
Mat descriptors;
bowDE.compute(img, keypoints, descriptors);
if (descriptors.empty()) return 0;
// setup svm
Ptr svm = SVM::create();
svm->load(svmLoad);
int prediction = svm->predict(descriptors);
cout << "Prdiction for " << queryLoad << " is: " << prediction << endl;
return prediction;
}
int main(int argc, char** argv)
{
//Deal with user input
//TODO: validate file paths
if (argc > 5 && (string)argv[1] == "/t" || (string)argv[1] == "/train")
{
TrainSVM(PathsToImageMats(GetFiles(argv[2])), PathsToImageMats(GetFiles(argv[3])), argv[4], argv[5]);
}
else if (argc < 6 && (string)argv[1] == "/p" || (string)argv[1] == "/predict")
Predict(argv[2], argv[3], argv[4]);
else
{
cout << "Accepted parameters are as follow:" << endl;
cout << "/t positive-folder-path negative-folder-path svm-save-path vocabulary-save-path" << endl;
cout << "/p query-image-path svm-load-path vocabulary-load-path" << endl;
}
system("pause");
}
↧