ACDNet: Adaptively Combined Dilated Convolution for Monocular Panorama Depth Estimation

Depth estimation is a crucial step for 3D reconstruction with panorama images in recent years. Panorama maintain the complete spatial information but introduce distortion equirectangular projection. In this paper, we propose an ACDNet based on adaptively combined dilated convolution to predict dense depth map monocular panoramic image. Specifically, combine kernels different dilations extend receptive field Meanwhile, adaptive channel-wise fusion module summarize feature maps and get diverse attention areas along channels. Due utilization of constructing module, network can capture leverage cross-channel contextual efficiently. Finally, conduct experiments three datasets (both virtual real-world) experimental results demonstrate that our proposed substantially outperforms current state-of-the-art (SOTA) methods. Our codes model parameters are accessed https://github.com/zcq15/ACDNet.