Paper Title
Optimizing Dynamic Occlusion in Mobile AR: From Vision Transformers to Guided Filter Pipelines

Abstract
Augmented Reality acts as a bridge between rendering of virtual objects to real world objects. Realistic occlusion is major part in Augmented reality in terms of achieving accurate rendering of virtual object to real world where virtual object is perfectly rendered behind or in front of real world object. But achieving realistic occlusion on android mobile phones still remains challenge due to lacking of dedicated LiDAR sensors for depth estimation. While we look back the evolution of MDE, we specifically address the boundary blurring in lightweight models by proposing an edge-aware depth refinement pipeline. We track the technologies used to solve this challenge such as early CNN based models to Vision Transformers and Depth Anything. But these models sometimes struggle with achieving speed on mobile devices while trading off accuracy. To balance this trade-off, we analyze the rising trend of Knowledge Distillation [1], a technique where massive ”Teacher” Deep Neural Network AI models train lightweight ”Student” networks suitable for android devices to be both fast and accurate. To directly address boundary blurring in these lightweight models, we propose an edge-aware depth refinement pipeline utilizing an RGB-Guided Filter, demonstrating sharpened occlusion boundaries with minimal latency overhead. Keywords - Augmented Reality, Monocular Depth Estimation, Mobile AI, Guided Filter, Depth Refinement