Dreamed about training, verifying and validating your QoE model on a million videos?

Abstract : Training, verification, and validation of objective prediction models require well-chosen test stimuli. The measured prediction performance depends largely on the congruence of stimulus selection in the three steps training, verification, and validation. Different stimulus selection criteria are discussed: extracting a representative set of stimuli from the scope of application, spreading the range of application scope with equidistant stimuli, or using stressful stimuli for the prediction algorithm. Nowadays, most databases are too small to sufficiently cover even one of these evaluation types; a large-scale database may solve the problem but requires new statistical methods and understanding of quality evaluation.