A Unified Multiple Inducible Co-attentions and Edge Guidance Network for Co-saliency Detection
ARTIFICIAL NEURAL NETWORKS AND MACHINE LEARNING - ICANN 2022, PT I(2022)
摘要
The learning-based methods have improved the performances of co-salient object detection (CoSOD). Mining the intra-image saliency individuals and exploring the inter-image co-attention are two challenges. In this paper, we propose a unified Multiple INducible co-attentions and Edge guidance network (MineNet) for CoSOD. Firstly, a classified inducible co-attention (CICA) is designed to model the classification interactions from a group of images. Secondly, a focal inducible co-attention (FICA) is employed to adaptively suppress and aggregate inter-image saliency features. CICA and FICA are jointly embedded into the network to predict the co-attention. The co-attentions of CICA and FICA are collaborative calibration and mutual optimization. Thirdly, we put forward an edge guidance module (EGM) to mine the intra-image saliency individuals, which aims to keep the consistency of co-attention during the feature transfer and refine the object edges. Finally, these three modules are merged into a unified and end-to-end network to predict the fine-grained boundary-preserving salient objects. Experimental results on three prevailing benchmarks show that our MineNet outperforms other competitors in terms of the evaluation metrics. In addition, the proposed method runs at the speed of more than 30 fps on a single GPU.
更多查看译文
关键词
Co-salient object detection,Co-attention,Edge guidance module
AI 理解论文
溯源树
样例
![](https://originalfileserver.aminer.cn/sys/aminer/pubs/mrt_preview.jpeg)
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要