Staff Profile Pages

ImmersiveDepth: A Hybrid Approach for Monocular Depth Estimation from 360 Images Using Tangent Projection and Multi-Model Integration

Authors: Dorosti, S. and Yang, X.

Journal: Proceedings 2025 IEEE Conference on Virtual Reality and 3D User Interfaces Abstracts and Workshops Vrw 2025

Pages: 1392-1393

DOI: 10.1109/VRW66409.2025.00342

Abstract:

ImmersiveDepth is a hybrid framework designed to tackle challenges in Monocular Depth Estimation (MDE) from 360-degree images, specifically spherical distortions, occlusions, and texture inconsistencies. By integrating tangent image projection, a combination of convolutional neural networks (CNNs) and transformer models, and a novel multi-scale alignment process, ImmersiveDepth achieves seamless and precise depth predictions. Evaluations on diverse datasets show an average 37% reduction in RMSE compared to Depth Anything V2 and a 25% accuracy boost in low-light conditions over MiDaS v3.1. ImmersiveDepth thus establishes a robust solution for immersive technologies, autonomous systems, and 3D reconstruction.

Source: Scopus