Patch-Wise Attention Network for Monocular Depth Estimation

In computer vision, monocular depth estimation is the problem of obtaining a high-quality map from two-dimensional image. This provides information on three-dimensional scene geometry, which necessary for various applications in academia and industry, such as robotics autonomous driving. Recent studies based convolutional neural networks achieved impressive results this task. However, most previous did not consider relationships between neighboring pixels local area scene. To overcome drawbacks existing methods, we propose patch-wise attention method focusing each area. After extracting patches an input feature map, our module generates maps patch, using two modules patch along channel spatial dimensions. Subsequently, return to their initial positions merge into one feature. Our straightforward but effective. The experimental challenging datasets, KITTI NYU Depth V2, demonstrate that proposed achieves significant performance. Furthermore, outperforms other state-of-the-art methods benchmark.