Improving Pedestrian Attribute Recognition with Multi-Scale Spatial Calibration

Pedestrian Attribute Recognition (PAR) has attracted increasing attention since it could provide important structural information of pedestrians for Smart Video Analysis. However, the pedestrian images are taken from a far distance significantly increase the difficulty of PAR for fine-grained attributes. To address these problems, and further improve the effects of PAR, we proposed a Multi-Scale Spatial Calibration (MSSC) module. More specifically, the module includes two submodules: first, a Spatial Calibrated Module (SCM) is proposed to extract more discriminative features of inconspicuous attributes from its surrounding regions by gathering the contextual information across different receptive fields. Moreover, in order to build the long-range dependencies of pyramid feature maps in different spatial scales, we also propose Multi-Scale Feature Fusion (MSFF) to integrate the multiple branches of low-level detailed features and high-level semantics features by non-local attention mechanism. Extensive experiments show that our proposed model could achieve state-of-the-art results on three pedestrian attribute datasets, including RAPv1, PA-100K, and RAPv2. Especially, the proposed model significantly improves the recognition effects of fine-grained attributes in low-resolution images in terms of mean Accuracy (mA) and recall. Code is available at https://github.com/iceicei/MSSC.