FsaNet: Frequency Self-Attention for Semantic Segmentation

Considering the spectral properties of images, we propose a new self-attention mechanism with highly reduced computational complexity, up to a linear rate. To better preserve edges while promoting similarity within objects, we propose individualized processes over different frequency bands. In particular, we study a case where the process is merely over low-frequency components. By ablation study, we show that low frequency self-attention can achieve very close or better performance relative to full frequency even without retraining the network. Accordingly, we design and embed novel plug-and-play modules to the head of a CNN network that we refer to as FsaNet. The frequency self-attention 1) requires only a few low frequency coefficients as input, 2) can be mathematically equivalent to spatial domain self-attention with linear structures, 3) simplifies token mapping (1×1 convolution) stage and token mixing stage simultaneously. We show that frequency self-attention requires 87.29% ~ 90.04% less memory, 96.13% ~ 98.07% less FLOPs, and 97.56% ~ 98.18% in run time than the regular self-attention. Compared to other ResNet101-based self-attention networks, FsaNet achieves a new state-of-the-art result (83.0% mIoU) on Cityscape test dataset and competitive results on ADE20k and VOCaug. FsaNet can also enhance MASK R-CNN for instance segmentation on COCO. In addition, utilizing the proposed module, Segformer can be boosted on a series of models with different scales, and Segformer-B5 can be improved even without retraining. Code is accessible at https://github.com/zfy-csu/FsaNet.