GRdepth: Enrich feature with global information and self-iterative regulation network for monocular depth estimation

Abstract

Monocular depth estimation (MDE) seeks to infer pixel-wise dense depth maps from a single RGB image. Recent methodologies predominantly utilize the encoder-decoder architecture to effectively extract and analyze multi-scale features. However, they tend to ignore the important role that high-level features with rich global information play in MDE, resulting in a poor understanding of the overall structure of the scene by the model. Based on this, we propose a novel encoder-decoder framework called GRdepth, which includes a cross large scale feature enhancement (CLSE) module and an iterative regulation decoder (IRD). Specifically, the CLSE module is designed to use high-level features, enriched with global information extracted by a global information aggregation (GIA) unit, to guide the enhancement of multi-scale feature maps produced by the encoder. This enhancement is achieved through a cross large scale feature fusion (CLSF) unit built from channel attention and spatial attention to refine low-level features with high-level information. The IRD is tailored for MDE based on classification-regression which mainly utilizes a bin width self-regulation (SRbins) unit to adjust the width of the initial bins predicted with the bottleneck features. This adjustment is guided by bin width predicted by an iterative adaptive feature fusion (IAFF) unit at each level, effectively combining global information and local information for more accurate bin width and bin centers. Extensive experiments on the indoor dataset NYU-Depth-v2 and SUN-RGBD and on the outdoor dataset KITTI demonstrate that our method can achieve comparable state-of-the-art (SOTA) results.