D2CNet: A Densely Connected Collaborative Network for Saliency Detection and Object Size Prediction

Abstract:
U-shape networks and feature pyramid networks (FPN) have shown their advantages in the saliency prediction task. However, most existing saliency methods usually fail to detect salient objects of different sizes. In this paper, we propose a Densely Connected Collaborative network (D2CNet) which considers both multi-scale input images and multi-level feature blocks to deal with the above problem. The proposed D2CNet consists of three modules: bottom-up encoder module, top-down decoder module and size prediction module. In the encoder module, we use two backbones with shared parameters to extract multi-level features from two inputs and these multi-level features must be transformed to generate more robust salient visual features. Next, in the decoder module, all the transformed feature blocks are gradually used to generate the final saliency map. Finally, fully connected layers and sigmoid function are used to predict object size. In addition, different from binary cross entropy, the proposed size-aware loss can guide the D2CNet to detect objects of different sizes under complex background. Experimental results on five popular saliency datasets demonstrate that D2CNet outperforms ten state-of-the-art methods on different evaluation metrics.