R-TAFM: purely convolutional recurrent refinement for deployment-oriented monocular depth estimation

Abstract Monocular depth estimation is widely used for navigation and scene understanding, yet deployment requires balancing accuracy with predictable runtime and compact models. Many recent lightweight designs pair depthwise separable convolutions with transformer components to boost accuracy, which typically introduces a more diverse operator set and can make realized throughput more dependent on the deployment backend in practice. Instead, we revisit recurrent refinement from a deployment-oriented perspective and introduce R-TAFM, purely convolutional framework that performs iterative depth refinement at a fixed working resolution with a parameter-shared decoder. We further derive a deployment-mode variant, R-TAFM-Fast, which is trained with recurrent supervision yet reduces inference to a single decoder pass, lowering latency on commodity GPU and on a Jetson-class embedded GPU. For self-supervised learning, we introduce an adaptive reprojection objective that jointly handles occlusions and independently moving objects without auxiliary tasks, and a neighborhood-consistent correction of auto-masked stationary pixels to prevent supervision collapse in homogeneous regions. Both quantitative benchmarks and qualitative assessments demonstrate that, with $$\approx $$ ≈ 3.2M parameters, R-TAFM achieves accuracy comparable to or exceeding recent lightweight state-of-the-art methods, using only standard convolutions.