Single image depth estimation using improved U-Net and edge-guide loss

Monocular depth estimation is regarded as a critical link in context-aware scene comprehension, which typically uses image data from a single point of view as the input to directly predict the depth value corresponding to each pixel in the image. However, predicting accurate object borders without replicating texture is difficult, resulting in missing tiny objects and blurry object edge in predicted depth images. In this paper, we propose a method for estimating monocular depth using an improved U-Net-based encoder-decoder network structure. We propose a new training loss term called edge-guide loss, which pushes the network to focus on object edges, resulting in better accuracy of the depth of tiny objects and edges. In the network, we build the encoder using DenseNet-169 and the decoder using 2 × bilinear up-sampling, skip-connections and hybrid dilated convolution. And skip-connections are used to send multi-scale feature maps from encoder to decoder. We specifically create a new loss function, edge-guide loss and three basic loss terms. We test our algorithm on the NYU Depth V2 dataset. The results of the experiments show that the proposed network can create depth image from a single RGB image with unambiguous borders and more tiny object depth. In the meantime, compared with state-of-the-art approaches, our proposed network outperforms for both visual quality and objective measurement. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.