Feature extraction is one of two core tasks of a person re-identification besides metric learning. Building an effective feature extractor is the common goal of any research in the field. In this work, we propose a deep spatio-temporal network model which consists of a VGG-16 as a spatial feature extractor and a GRU network as an image sequence descriptor. Two temporal pooling techniques are investigated to produce compact yet discriminative sequence-level representation from a sequence of arbitrary length. To highlight the effectiveness of the final sequence-level feature set, we use a cosine distance metric learning to find an accurate probe-gallery pair. Experimental results on the ilIDS-VID and PRID 2011 dataset show that our method is slightly better on one dataset and significantly better on the other than state-of-the-art ones.