A Monocular Depth Estimation Method for Indoor-Outdoor Scenes Based on Vision Transformer

Abstract

In the field of computer vision, monocular depth estimation has garnered significant attention as a research direction. However, current depth estimation methods often overlook the impact of depth range variations in indoor and outdoor scenes, consequently limiting the model's generalization ability. To achieve high-precision depth estimation across different depth ranges, we propose a new method. We employ the pretrained model Dinov2 as encoder, combined with decoder based on CNN architecture, to enhance the network's capacity for extracting global information from indoor and outdoor scenes. Also, we design a mapping module to transform diverse depth ranges into a unified 0-1 range, which can effectively adapt to indoor and outdoor scenes. We validate our method on the DIODE dataset, which comprises mixed indoor and outdoor scenes. Experimental results demonstrate that our method achieves higher depth estimation accuracy and stronger generalization performance when dealing with scenes of diverse depth ranges.