TransIndoor: Transformer Based Self-supervised Indoor Depth Estimation

In various tasks of self-supervised monocular depth estimation, the prevalent norm has been the adoption of a U-shaped network structure, utilizing the convolution module as the foundational operator, leading to notable accomplishments. However, the inherent limitation of convolution operations, characterized by a restricted receptive field, often poses challenges in explicitly capturing long-range dependencies. The Transformer, originally designed for sequence-to-sequence prediction, presents a global self-attention mechanism capable of capturing long-range dependencies but may compromise localization abilities, lacking in low-level details. This paper introduces TransIndoor, a robust alternative for self-supervised monocular depth estimation that combines the strengths of the Transformer and convolution. TransIndoor effectively extracts both global context information and local spatial details simultaneously. Furthermore, a novel local multi-scale fusion block is introduced to enhance fine-grained details by processing skipped connections within the encoder through the primary CNN stem. The comprehensive validation of TransIndoor using the NYU Depth V2 dataset and ScanNet demonstrates its capability to generate satisfactory depth maps, addressing the limitations of existing methods.