Optimising Real-Time Stereo Image Retargeting for AR/VR: Lightweight Disparity CNNs on AI-Driven Edge Architectures

Real-time stereo image retargeting for augmented reality (AR) and virtual reality (VR) necessitates precise per-pixel depth estimation and ultra-low latency performance on resource-limited edge devices. Current disparity convolutional neural networks (CNNs) and retargeting pipelines are unable to meet these demanding requirements concurrently. This paper presents EASNet, a compact end-to-end framework that unifies geometry-aware proposal generation, parallax-aligned feature encoding, sparse candidate aggregation, and uncertainty-guided refinement to enable high-fidelity stereo retargeting on edge architectures. This system enhances stereo vision through Epipolar-Adaptive Disparity Proposals (EADP) for search space reduction, a Parallax- Directed Deformable Encoder (PaDDE) for improved matching in repetitive and low-texture areas, Sparse Epipolar Candidate Volume (SECV) with Edge-Consistent Routing (ECR) for efficient, boundary-preserving cost aggregation, and Lightweight Uncertainty-Guided Refinement (LUGR) for subpixel structure and occlusion correction. Evaluated on high-resolution indoor stereo data, EASNet attains a favourable trade-off between accuracy and efficiency (≈0.21M parameters, 3.42 GFLOPs) while improving disparity fidelity and visual coherence required for retargeting (reported EPE ≈ 1.78 px, D1 ≈ 5.02%, VC ≈ 93.7%). The design emphasises quantisation compatibility and deterministic latency, enabling practical deployment on AR/VR edge devices. We analyse ablations, per-scene behaviour, and k-fold stability, and we discuss limitations, notably the indoor training bias and remaining challenges in extreme occlusion and large baseline scenarios, together with directions for extending EASNet via multi-scale context fusion and crossdomain adaptation.