SwinFusion: Channel Query-Response Based Feature Fusion for Monocular Depth Estimation

Given a single RGB image, Monocular Depth Estimation (MDE) predicts the depth of each pixel by utilizing both global and local information. Most MDE networks use simple skip connections, each skip connection only contain single-scale feature information. In this paper, we propose a multi-scale feature fusion module called Channel Query-Response Fusion (CQRF), which uses a channel query-response mechanism to filter out harmful or invalid channel information and enables soft fusion on the channels of multi-scale features. It can be inserted as an independent module into other MDE network architectures and enhance their performance. After passing through the CQRF module, the skip connections will contain valid channel information from all scales. Based on CQRF, we design a multi-scale feature fusion network SwinFusion, which achieves advanced results on public dataset. Additionally, we analyzed the effects of the number of utilized multi-scale features and different layer channels on the performance of MDE task.