Dual-Stream Multiscale Attention Monocular Depth Estimation Network

Abstract

Deep-learning-based methods have shown superior performance in monocular depth estimation tasks. However, the existing methods often overlook small-scale objects and vertical information while suffering from edge blurring and loss of low-texture information issues. To remedy the issue, a dual-stream multiscale attention network (DMA-Net) for monocular depth estimation is proposed, featuring an encoder-decoder pattern. Specifically, two scales of inputs are, respectively, fed into the pretrained ResNeXt-101 to extract diverse image features. Then, a multiscale attention feature fusion model is constructed, where self-attention dilated convolution blocks effectively capture multiscale global features with long-distance dependencies and feature fusion blocks promote information exchange between the two tributaries, further reinforcing features. Next, a guiding decoder is designed to refine the restored depth map by assistively integrating the outputs of each encoder layer, and exploit efficient channel attention network to recalibrate the meaningful information. Finally, vertical information extractor is utilized to capture vertical features for enhancing the restore ability of longitudinal depth details. Extensive experiments are conducted on the KITTI and NYU Depth V2 datasets, and the results show that the proposed DMA-Net outperforms all previous methods, achieving competitive results on the majority of the metrics.