MEMF: Multi-level-attention embedding and multi-layer-feature fusion model for person re-identification

Abstract

Person re-identification (re-ID) methods need to extract representative, rich and discriminative features in order to deal with the effect of imperfect pedestrian detectors, illumination changes, occlusions, and background confusion. In this paper, a multi-level-attention embedding and multi-layer-feature fusion (MEMF) model is proposed for person re-ID. Specifically, a novel backbone network is designed, in which multi-level-attention blocks are embedded into a multi-layer-feature fusion architecture. Multi-level-attention blocks can highlight representative features and assist global feature expression, and multi-layer-feature fusion can increase the fine granularity of feature expression and obtain richer features. Besides, a new eigenvalue difference orthogonality (EDO) loss is designed to reduce the correlation between features. The final loss is defined as the combination of the cross-entropy loss and the EDO loss, which improves re-ID results. The proposed method is evaluated on four popular and challenging datasets. Detailed experiments demonstrate that the application of various elements of the MEMF model can help improve person re-ID performance. Compared with start-of-the-art methods, the MEMF model gets a promising result.