Iterative contextual and adaptive strategies for enhanced monocular depth estimation

초록

Monocular Depth Estimation (MDE) is an important computer vision task within artificial intelligence (AI). It is widely used in fields such as autonomous navigation, robotics, and augmented and virtual reality. Current approaches often face difficulties in accurately estimating depth, particularly in scenes with complicated shapes, occlusions, and diverse depths. To overcome these challenges, we introduce an innovative AI framework that combines two main methods: Iterative Elastic Bins (IEBins) and a Contextual Transformer-based Feature Extractor (CTFE). The IEBins method improves depth predictions by repeatedly refining the depth estimation bins according to the scene's depth uncertainty. The CTFE method, using a Swin Transformer architecture, extracts both local and global context information, which significantly helps refine depth predictions. Our integrated approach, called the Progressive Refinement Strategy, effectively reduces errors like blurred edges and inconsistent depth predictions. Comprehensive testing on standard datasets—including Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI), New York University Depth Dataset Version 2 (NYU-Depth-v2), and Scene Understanding RGB-Dataset (SUN RGB-D)—demonstrates that our method outperforms existing models in accuracy and efficiency. Moreover, our AI model generalizes effectively to new, unseen scenarios, and is computationally efficient enough for real-time use in practical applications such as autonomous vehicles, robotics systems, and immersive virtual and augmented reality experiences. © 2025 Elsevier Ltd

키워드