Embedded Attention Network for Semantic Segmentation

Semantic segmentation, as a fundamental task in computer vision, is capable of providing perception ability in many robot applications, such as automatic navigation. To enhance the segmentation accuracy, self-attention mechanism is adopted as a key technique for capturing the long-range dependency and enlarging the receptive fields. However, it requires high computation complexity and GPU memory. In this letter, we propose an Embedded Attention Network to relieve the undesired computational cost. Specifically, we introduce an Embedded Attention (EA) block to improve the segmentation performance and efficiency. Firstly, EA block generates a group of compact while coarse feature bases with the capability of reducing large amount of computation cost. Then an embedded attention is employed to collect the global contextual information and update the representation of the coarse bases from a global view. Finally, the updated bases are leveraged to estimate the attention similarity. We take the well-estimated feature bases to perform feature aggregation. Our approach achieves a considerable computation cost reduction, which suggests it is more powerful than other counterparts in most robot platforms. We conduct extensive experiments on two benchmark semantic segmentation datasets, i.e., CityScapes and ADE20 K. The results demonstrate that the proposed Embedded Attention network delivers comparable performance with high efficiency.