SA-DETR: Saliency Attention-based DETR for salient object detection

Researches on the Salient Object Detection (SOD) task have made many advances based on deep learning methods. However, most methods have focused on predicting a fine mask rather than finding the most salient objects. Most datasets for the SOD task also focus on evaluating pixel-wise accuracy rather than “saliency”. In this study, we used the Salient Objects in Clutter (SOC) dataset to conduct research that focuses more on the saliency of objects. We propose a architecture that extends the cross-attention mechanism of Transformer to the DETR architecture to learn the relationship between the global image semantics and the objects. We extended module with Saliency Attention (SA) to the network, namely SA-DETR, to detect salient objects based on object-level saliency. Our proposed method with cross- and saliency-attentions shows superior results in detecting salient objects among multiple objects compared to other methods. We demonstrate the effectiveness of our proposed method by showing that it outperforms the state-of-the-art performance of the existing SOD method by 4.7% and 0.2% in MAE and mean E-measure, respectively.