Pixel-level Monocular Depth Estimation Based on Global-local Feature Fusion

To address the issue of insufficient extraction of global features in CNN monocular depth estimation networks, leading to scale ambiguity and object edge ambiguity in predicted depth maps, this paper presents a pixel-level monocular depth estimation method based on global-local feature fusion. Multiple pre-trained Transformer blocks are employed as encoders to extract global features at multiple scales. Additionally, a global area enhancement module based on pyramid pool is designed to increase the receptive field of the network, providing effective global context prior for in-depth scene analysis and enhancing the network's ability to obtain feature information. Furthermore, a feature fusion module based on large kernel convolution attention is proposed to better utilize global and local features such as shape and texture, refining semantic scenes and improving the global contours of depth maps. Finally, an adaptive distribution depth interval and estimation module are designed according to the requirements of depth features in different regions of the scene. The depth information is estimated by the probability distribution of pixels and linear combination of discrete pixels, resulting in clearer predicted depth maps with detailed and smooth object outlines.