1. Open set recognition: Unknown individuals may appear on scene (CCTV).
  2. Gallery = 10 individuals with a mug-shot image per individual.

Approach: Gaussian Mixture Model-UBM based

  1. Use video sequences of 20 individuals of random people that might not appear in the gallery or in the scene as UBM (universal background model).

    ubm = gmm_em(datalist, numberOfMixtures, EMiterations, downSamplingfactor, parallelWorker)

'datalist' contain feature vector of each individuals that assigned as UBM (20xnumberOfvideoSequences) cells. Each cell contains [numberOfMixtures x numberOfFrames] double.

  1. Adapt the ubm into each individuals (in gallery) to create GMM speaker model.

    for I = 1:numberOfIndividuals gmm{s} = mapAdapt(spkFiles(I,:),ubm, map_tau, config) end

spkFiles contain feature vector of the mugshot images for each individuals in the gallery. (10 x 1) cells, which each cell contain [numberOfMixtures x 1] double.

The approach above is based on MSR toolkit for speaker verification. However, I'm not so sure about implementing it for image verification. Is it what I had described above is correct? The way to prepare the datalist and spkFiles? How to decide the number of mixtures? Can I use any feature extraction method such as PCA?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.