How do I train an LBPHFaceRecognizer Model with a very large number of images?

C++
,
1

How do I train an OpenCV LBPHFaceRecognizer Model with a very large number of images?

I’m trying to create a model in OpenCV for face detection (specifically sex and age). The data set I have has about 50,000 images. The problem is; when I try to train a model with that amount of data, the memory usage just grows and runs out of memory after about 100 images. Even when trying batching and unloading the loaded cv::mat image, the model just grows in size (which I guess is to be expected).

I tried both loading and training one image at a time, and with this variant I tried training in batches but still the same result

The ImageCollection class is just a map of SEXAGEBRACKETs (labels) to image file paths. The model I’m using is LBPHFaceRecognizer.

Here is what my “load and train” function looks like:

void ModelGenerator::LoadImagesAndTrain(const ImageCollection* imageCollection, const cv::Ptr<cv::face::LBPHFaceRecognizer> modelToTrain, const bool isUpdatingModel)
    {
    	cv::Size loadedImageSize;
    	cv::Mat currLoadedImage;
    	std::vector<cv::Mat> imagesToUse;
    	std::vector<int> labelsToUse;
    	std::multimap<SEXAGEBRACKET, char*>* bracketToImageFileNameMap;
    
    	if (imageCollection != NULL && modelToTrain != NULL)
    	{
    		std::cout << "Number of images provided for training:" << imageCollection->GetNumberOfImages() << std::endl;
    
    		bracketToImageFileNameMap = imageCollection->GetBracketToFileNameMap();
    
    		//Check if this is the first time we are training the model, All other calls will be an update instead.
    		if (isUpdatingModel == false)
    		{
    			currLoadedImage = cv::imread(bracketToImageFileNameMap->begin()->second, cv::IMREAD_GRAYSCALE);
    
    			imagesToUse.push_back(currLoadedImage);
    			labelsToUse.push_back(bracketToImageFileNameMap->begin()->first);
    
    			modelToTrain->train(imagesToUse, labelsToUse);
    
    			currLoadedImage.release();
    		}
    
    		imagesToUse.clear();
    		labelsToUse.clear();
    
    		for (std::multimap<SEXAGEBRACKET, char*>::iterator it = bracketToImageFileNameMap->begin(); it != bracketToImageFileNameMap->end(); it++)
    		{
    			std::cout << "training model...." << std::endl;
    
    			imagesToUse.clear();
    			labelsToUse.clear();
    
    			std::cout << it->first << ":" << it->second << std::endl;
    
    			currLoadedImage = cv::imread(it->second, cv::IMREAD_GRAYSCALE);
    
    			imagesToUse.push_back(currLoadedImage);
    			labelsToUse.push_back(it->first);
    
    			modelToTrain->update(imagesToUse, labelsToUse);
    			
    			currLoadedImage.release();
    		}
    	}
    }

Is it possible to create a very large model (with something like 50,000 images) in OpenCV without needing like 300GB of Ram? Or is there an alternative to doing with with a Neural Network? I’m open to suggestions if I’m going about this the wrong way.

2

I tried both loading and training one image at a time, and with this variant I tried training in batches but still the same result

that’s the obvious, expected result.
batches or not, each train image will add a ~64k row of features to your database

Is it possible to create a very large model

LBPH does not build a ‘model’, it’s only a collection of (linearly independant !) features. unlike pca or ann based models, you don’t need feature A and B to derive feature C.
in other words, HERE you can have 10 small databases (and just collect the response with smallest distance from many db’s) instead of accumulating a 300gb chunk

I’m trying to create a model in OpenCV for face detection (specifically sex and age)

ouch. detection ? (this is a classification process)
are you aware, there are cnn based models available for this in opencv since ~10 years ?

just a map of SEXAGEBRACKETs (labels) to image file paths.

do you have a link for us ?

3

I’m not sure where I should go from here, should I not be using LBPH to try and categorize Age. Can you tell me more about the CNN models? I haven’t worked with CNNs before so I’m not sure where to proceed or maby use a different model? or perhaps modify the images somewhat?