DepthVision3D Dataset

Deep learning-based 3D reconstruction (also known as depth estimation in some applications) from one or a pair of images has become an important area of engineering applications. However, its research progress has been constrained by the lack of high-quality training datasets. Here, we have prepared a new high-accuracy versatile dataset designed for such deep learning-based 3D reconstruction and depth estimation methods. The dataset provides rich images and labels for a variety of supervised and unsupervised learning. To show the usage of the dataset, we also present a Hierarchical Depth Network (HiDNet), which is a generalized image-to-depth regression network implemented in PyTorch. HiDNet shows superior performance over the standard UNet architecture by retaining more spatial information in the network layers and using the attention mechanism to capture important information, making it a starting point and effective baseline for future research and development in this field.