Transformer-based hierarchical dynamic decoders for salient object detection

Qingping Zheng, Ling Zheng, Jiankang Deng, Ying Li, Changjing Shang, Qiang Shen

Abstract: Highlights•T-HDDNet employs dynamic upsampling and fusion for data-driven salient object detection.•T-HDDNet uses a spatially sensitive loss to learn salient object structures, yielding a high-resolution map.•Our method exceeds benchmarks at 22.19G MACs and 27M Params, reducing costs compared to the top model.