Doubly-Fused ViT: Fuse Information from Vision Transformer Doubly with Local Representation.

Vision Transformer (ViT) has recently emerged as a new paradigm for computer vision tasks, but is not as efficient as convolutional neural networks (CNN). In this paper, we propose an efficient ViT architecture, named Doubly-Fused ViT (DFvT), where we feed low-resolution feature maps to self-attention (SA) to achieve larger context with efficiency (by moving downsampling prior to SA), and enhance it with fine-detailed spatial information. SA is a powerful mechanism that extracts rich context information, thus could and should operate at a low spatial resolution. To make up for the loss of details, convolutions are fused into the main ViT pipeline, without incurring high computational costs. In particular, a Context Module (CM), consisting of fused downsampling operator and subsequent SA, is introduced to effectively capture global features with high efficiency. A Spatial Module (SM) is proposed to preserve fine-grained spatial information. To fuse the heterogeneous features, we specially design a Dual AtteNtion Enhancement (DANE) module to selectively fuse low-level and high-level features. Experiments demonstrate that DFvT achieves state-of-the-art accuracy with much higher efficiency across a spectrum of different model sizes. Ablation study validates the effectiveness of our designed components.