Lightweight person re-identification model employing symmetrical combination units

Abstract

As an image retrieval problem, person re-identification (Re-ID) relies on robust features extracted by convolution neural models. Most current methods use large backbone models for feature extraction (e.g., ResNet50). However, these large backbone models have many parameters, which cause many problems when embedded in smart camera devices. For example, the device's computing resources are limited, the real-time operation speed is limited, etc. So it is necessary to construct models with low parameters and low complexity. This paper proposes a new lightweight baseline for Re-ID, which is SCL-net and all underlying modules of the model are reconstructed. In our work, we design a new convolution unit----symmetrical combination units (SC-unit), which construct features map of richer channels by reusing feature maps from different convolution layers. In addition, we redesigned all the base modules of SCL-net and proved the effectiveness of all modules. We joint training of shallow and deep features of the model respectively to improve the accuracy of the model. Our SCL-net has about 2.3M parameters, and it can achieve 95.2%/85.9% on Rank-1 and mAP without any pretraining.