SliceFormer: deep dense depth estimation from a single indoor omnidirectional image using a slice-based transformer

Abstract In this research, we tackle the task of estimating depth from a single indoor omnidirectional image, presented in an equirectangular projection. This problem arises from explaining the underlying principles of depth estimation in indoor environments and the contributing factors. Mainstream depth estimation methods, which predominantly utilise deep learning, exhibit limitations due to their heavy reliance on data-driven patterns, often overlooking underlying principles. Recognising this gap, our experimental evaluation substantiates the significant impact of gravity alignment for depth estimation. Building on this foundation, we process the input equirectangular projection by dividing it into vertical slices aligned with the gravity direction. These slices are then utilised as patch embeddings for a transformer encoder to predict an equirectangular depth map. Our architecture is evaluated against state-of-the-art models using real-world datasets, namely Stanford2D3D and Matterport3D, demonstrating superior performance. These results underscore the significance of our gravity-aligned approach for depth estimation in omnidirectional images. The project is publicly available at https://github.com/MinisculeDust/SliceFormer .