A sample selection method and apparatus and a server belong to the field of metric learning technologies. The method includes: selecting n sample pairs from an unlabeled sample set, each sample pair including two samples, and each sample including data in p modalities; calculating a partial similarity between data that is in each modality and that is of one sample included in the sample pair and data that is in each modality and that is of the other sample, to obtain p×p partial similarities; calculating, according to the p×p partial similarities, an overall similarity between the two samples included in the sample pair; obtaining a degree of difference between the p×p partial similarities and the overall similarity; and selecting a sample pair that meets a preset condition and that is in the n sample pairs as a training sample. In this application, training samples of high quality are selected to train a metric model, so that the metric model of higher precision can be trained by using fewer training samples.