Attention U-Net Oriented Towards 3D Depth Estimation

Advancements in 3D reconstruction, applied to depth estimation in both indoor and outdoor environments, have achieved significant performance in various applications. Typically, outdoor environment reconstruction has relied on traditional approaches such as Structure from Motion (SFM) and its variants. In contrast, indoor environment reconstruction has shifted towards the use of depth-sensing devices. However, these devices have exhibited limitations due to environmental factors, such as lighting conditions. Recent approaches and optimizations have led to the development of novel methods that leverage Convolutional Neural Networks (CNNs), irrespective of whether the environment is enclosed or open. These methods are capable of complementing both approaches. In light of these advancements, alternatives have emerged, including the integration of Attention Layers. These have seen substantial proposals that have evolved over recent years. Thus, this paper proposes a method for 3D depth estimation with a focus on indoor and outdoor images. The goal is to generate high-detail and precise depth maps of real-world scenes using a modified U-Net network with the inclusion of a custom attention mechanism.