Two stream deep CNN-RNN attentive pooling architecture for video-based person re-identification

Abstract

Person re-identification (re-ID), is the task of associating the relationship among the images of a person captured from different cameras with non-overlapping field of view. Fundamental and yet an open issue in re-ID is extraction of powerful features in low resolution surveillance videos. In order to solve this, a novel Two Stream Convolutional Recurrent model with Attentive pooling mechanism is presented for person re-ID in videos. Each stream of the model is a Siamese network which is aimed at extracting and matching most differentiated feature maps. Attentive pooling is used to select most informative video frames. The output of two streams is fused to formulate one combined feature map, which helps to deal with major challenges of re-ID e.g. pose and illumination variation, clutter background and occlusion. The proposed technique is evaluated on three challenging datasets: MARS, PRID-2011 and iLIDS-VID. Experimental evaluation shows that the proposed technique performs better than existing state-of-the-art supervised video based person re-ID models. The implementation is available at https://github.com/re-identification/Person_RE-ID.git.