HiMoDepth: Efficient Training-Free High-Resolution On-Device Depth Perception

Depth perception is fundamental for various mobile and edge applications such as augmented reality, autonomous navigation, and computational photography. However, achieving high-resolution depth estimation on resource-constrained devices remains challenging due to computational and memory limitations. This paper presents HiMoDepth, an efficient training-free approach for high-resolution on-device depth perception. Instead of relying on expensive deep learning models that require extensive training data, HiMoDepth leverages geometric constraints and lightweight optimization techniques to estimate depth from monocular images. The system adaptively adjusts its computational complexity based on the device capabilities and application requirements while maintaining high-quality depth estimation. Comprehensive evaluations on various mobile devices and diverse scenes demonstrate that HiMoDepth achieves comparable accuracy to state-of-the-art learning-based methods while requiring significantly less computational resources and no training data.