RHN: RoI Restricted Hybrid Network for Instance-Aware Image-to-Image Translation

Image-to-image translation recently attracts a lot of attention, while it also encounters challenges when processing images which contain numerous instances. In this letter, a novel RoI restricted Hybrid Network (RHN) is proposed for instance-aware image-to-image translation. Firstly, we propose a batch-wise RoI restricted module which combines region restriction with batch-wise attention to recalibrate instance-level features. Secondly, we design a Transformer block with local Xception which can supplement local information on global features extracted from attention, to provide discriminative multi-scale features. Finally, we formulate a multi-task loss which can enforce RHN to achieve good instance-ware image-to-image translation performance. Experiments on publicly available datasets demonstrate the effectiveness of RHN.