VideoToolsLib

VideoToolsLib is an attempt to follow the DRY (don't reapeat yourself) principle. It'll be a collection of developt independent code components that I find useful from projec to project. It'll usually be some ML-algorithms purposed for CV (computer vision) usage, and it will heavly rely on OpenCV.

Compress data

So, you have a bunch of data samples and a need to compress them into something a ML algorithm can use. PCA is the thing for you! So, how? Generic interface:

class CCompressorInterface
{
public:
    virtual cv::Mat compress(cv::Mat _sample) = 0;
    virtual cv::Mat decompress(cv::Mat _compressedSample, bool _convertTo8Bit = false) = 0;
};

and a CPcaCompression class which follows that. Simple steps:

	CPcaCompression c(sampleNumber, sampleSize, dimensionsToUse);
	// train PCA
	while(  a_lot_of_samples ){
		c.addSample(smallSample);
	}
	c.process();
	// main usage
	cv::Mat smallData = c.compress(bigData);
	cv::Mat bigData = c.decompress(smallData); // simple, no?

NMS

A bunch of detection in the format of pairs std::pair<cv::Rect, double> and you need to NMS it? No problem.

	tVecDetections input;
	input.push_back(tDetectionPair(Rect(12,12,2,2), 0.2));
	...
	CFilterNMS filter(0.8);
	filter.process(input);

Voila!

Input transformations

There should be a bunch of transformer classes that you can chain-up to get the desired input. Currently:

• crop
• grayscale conversion
• image size normalization

Which can be used as such:

    IXfmr pipeStart;
    pipeStart.add(new CXNormSize(Size(100, 100))).add(new CXGrayscale());
    Mat outSample = pipeStart.transform(inSample);

Dependencies

• OpenCV (used 2.4)
• glog (google logging library)
• qt (for testing and qmake)
• boost

Contact

firstname.lastname@epfl.ch