IFDepth: Iterative fusion network for multi-frame self-supervised monocular depth estimation

Abstract

Self-supervised monocular depth estimation has gained prominence due to its training efficiency and applicability in autonomous systems. However, existing methods often exhibit limitations in preserving depth relationships in texture-homogeneous scenes and recovering fine-grained structural details. We present IFDepth, an iterative multi-frame depth prediction framework that refines coarse depth estimates through synergistic integration of optical flow features and multi-scale contextual information. Our architecture introduces three key components: (1) a Motion Feature Encoder (MFE) for spatiotemporal motion pattern extraction, (2) a Feature-Depth Cross Attention Layer (FCAL) enabling cross-modal feature interaction, and (3) a Gated Recurrent Unit (GRU)-based refinement module that progressively enhances predictions without computationally expensive 3D volume operations. Through iterative feature fusion, IFDepth effectively recovers occluded regions and high-frequency details while maintaining geometrically consistent depth ordering. Extensive experiments on KITTI, Cityscapes, and Robotcar datasets demonstrate state-of-the-art performance, particularly in preserving scene details and accurate depth ordering, outperforming existing monocular training approaches.