A Co-saliency Detection Method Based on Attention Model
Jin Gu, Hongxu Jiang · 2019
Existing co-saliency detection algorithms mainly focus on extracting low-level features, failing to deal with variations in background scenes and foreground objects. Thus, we propose a weakly supervised co-classification method based on K-means clustering to decompose the complicated high-level co-salient feature extraction into identifying the common object class(es) and discovering the corresponding neural attention maps. By our approach, the high-level consistency can be mapped to class excitation maps (CEMs), which can capture the common foreground regions. Then we integrate CEMs with bottom-up saliency maps (BUSMs), detecting co-saliency based on both high-level and low-level features combinedly. Finally, we apply a fully connected Conditional Random Field (FC-CRF) model for accurate boundary recovery. Our method is novel in weakly supervised learning and combining high-level and low-level cues. Experiment results show that our method achieves the state-of-the-art results on three benchmark datasets.