Real-time lightweight self-supervised monocular depth estimation

Monocular depth estimation aims to predict depth information within a scene from a single RGB image, but many models remain computationally intensive for real-time inference on resource-constrained edge devices. This paper presents a lightweight self-supervised monocular depth estimation network that balances accuracy and efficiency through targeted encoder–decoder design. The encoder employed a synergistic modeling approach combining decomposable large-kernel convolutions and local depthwise convolutions to capture both long-range context and local details with low computational overhead. The decoder utilized cross-scale feature differences as guidance to dynamically fuse multi-scale features, enhancing detail recovery and geometric consistency under lightweight constraints. In addition, a temporal soft fusion reprojection loss was employed to better leverage the complementary information of forward and backward frames, improving the robustness of self-supervised training. The model contained 3.0 M parameters and required 3.5 GFLOPs of computation. On KITTI, it achieves Abs Rel=0.105 and δ₁=0.892. On Make3D, it achieves Abs Rel=0.308 in a zero-shot setting. On a Rockchip RK3588S, a hybrid-quantized multi-thread implementation runs at 67 frames/s. The results demonstrated that the proposed method achieved a favorable accuracy–efficiency balance on edge devices, making it suitable for real-time monocular depth estimation tasks.