Wei et al. [wei2014cnn] and Yang et al. [yang2016exploit] addressed this problem by devising a 2-stage pipeline for multi-labelling in which the model generates image patches first and then labels them. However, these methods overemphasize the generated patches, thereby neglecting surrounding context infomation and label dependencies. The idea of object localization is similar to the attention mechanism that has been successfully applied in many vision tasks [zhu2017learning, Guo_2019_CVPR, wen2020multilabel, you2020crossmodality, 8682335]. Figure 1 illustrates the importance of label dependencies, spatial and context information. Additionally, context has been demonstrated useful in various visual processing tasks, such as recognition and detection [Zhang_2020_CVPR].