Densely activated self-attention for semantic segmentation

Liwen Xiao, Wenze Liu, Zhicheng Wang, Yiran Wang, Zhiyu Pan, Hao Lu, Zhiguo Cao

Abstract: Highlights•Transformer-based segmentation has ”region unevenness,” leading to incomplete predictions.•We introduce Densely Activated Self-Attention to tackle ”region unevenness”.•We introduce DAM, which enables even features and complete predictions.