A half-precision compressive sensing framework for end-to-end person re-identification

Abstract

Compressive sensing (CS) approaches are useful for end-to-end person re-identification (Re-ID) in reducing the overheads of transmitting and storing video frames in distributed multi-camera systems. However, the reconstruction quality degrades appreciably as the measurement rate decreases for existing CS methods. To address this problem, we propose a half-precision CS framework for end-to-end person Re-ID named HCS4ReID, which efficiently recoveries detailed features of the person-of-interest regions in video frames. HCS4ReID supports half-precision CS sampling, transmitting and storing CS measurements with half-precision floats, and CS reconstruction with two measurement rates. Extensive experiments implemented on the PRW dataset indicate that the proposed HCS4ReID achieves 1.55 $$\times$$ speedups over the single-precision counterpart on average for the CS sampling on an Intel HD Graphics 530, and only half-network bandwidth and storage space are needed to transmit and store the generated CS measurements. Comprehensive evaluations demonstrate that the proposed HCS4ReID is a scalable and portable CS framework with two measurement rates, and suitable for end-to-end person Re-ID. Especially, it achieves the comparable performance on the reconstructed PRW dataset against CS reconstruction with single-precision floats and a single measurement rate.