SegCFT: Context-aware Fourier Transform for efficient semantic segmentation

Yinqi Zhang, Lingfu Jiang, Fuhai Chen, Jiao Xie, Baochang Zhang, Gaoqi He, Shaohui Lin

Abstract: Highlights•Context-aware Fourier Transform is proposed to replace self-attention for segmentation.•Hierarchical Fourier Transform fuses the token-channel contexts to reduce computation cost.•Adaptive Modulation Unit fuses the spatial-frequency features for performance improvement.•SegCFT achieves the best trade-off between segmentation and inference speed.