TY - JOUR
AU - Tiancun Guo
AU - Qiang Zhou
AU - Mingliang Gao
AU - Gwanggil Jeon
AU - David Camacho
AB - In recent years, with the advancement of deep learning, person re-identification (Re-ID) has become increasingly significant. The existing person Re-ID methods primarily focus on optimizing network architecture to enhance Re-ID task performance. However, these methods often overlook the importance of valuable features in distinguishing Re-ID tasks, leading to reduced model efficacy in complex scenarios. As a solution, we utilize the attention mechanism to develop the lightweight multiscale Attentional Squeeze-and-Excitation Network (MASENet) that can distinguish between significant and non-significant features. Specifically, we utilize the SEAttention (SE) module to amplify important feature channels and suppress redundant ones. Additionally, the Spatial Group Enhance (SGE) module is introduced to enable networks to enhance semantic learning expression and suppress potential noise autonomously. We conduct comprehensive experiments on Market1501, MSMT17, and VeRi-776 datasets and cross-domain experiments on MSMT17 Ñ Market1501 to validate the model performance. Experimental results prove that the proposed MASENet achieves competitive performance across all experiments.
IS - In press
M1 - In press
N2 - In recent years, with the advancement of deep learning, person re-identification (Re-ID) has become increasingly significant. The existing person Re-ID methods primarily focus on optimizing network architecture to enhance Re-ID task performance. However, these methods often overlook the importance of valuable features in distinguishing Re-ID tasks, leading to reduced model efficacy in complex scenarios. As a solution, we utilize the attention mechanism to develop the lightweight multiscale Attentional Squeeze-and-Excitation Network (MASENet) that can distinguish between significant and non-significant features. Specifically, we utilize the SEAttention (SE) module to amplify important feature channels and suppress redundant ones. Additionally, the Spatial Group Enhance (SGE) module is introduced to enable networks to enhance semantic learning expression and suppress potential noise autonomously. We conduct comprehensive experiments on Market1501, MSMT17, and VeRi-776 datasets and cross-domain experiments on MSMT17 Ñ Market1501 to validate the model performance. Experimental results prove that the proposed MASENet achieves competitive performance across all experiments.
PY - 9998
SE - 1
SP - 1
EP - 8
T2 - International Journal of Interactive Multimedia and Artificial Intelligence
TI - Multiscale Attentional Squeeze-And-Excitation Network for Person Re-Identification
UR - https://www.ijimai.org/journal/bibcite/reference/3526
VL - In press
SN - 1989-1660
ER -