Batch-Incremental Triplet Sampling for Training Triplet Networks Using Bayesian Updating Theorem