Leveraging unlabeled data via semi-supervised contrastive learning for enhanced monocular depth estimation

Accurate monocular depth estimation with limited annotations remains a critical challenge in real-world scenarios. Existing supervised approaches heavily rely on large-scale dense annotations, the acquisition of which is costly and difficult to scale, while self-supervised methods often depend on restrictive geometric assumptions and additional information such as camera parameters or multi-view consistency, thereby limiting their generalization under sparse or noisy supervision. To bridge this gap, we propose TriConDepth, a novel semi-supervised framework that leverages contrastive learning (CL) to explore both labeled and unlabeled data effectively. The method comprises three stages: unsupervised image-level CL to extract global geometry-invariant structures from unlabeled images, empowering the embedding space with global contextual awareness; supervised pixel-level CL performed on labeled images to enhance local depth-aware feature discriminability; and a consistency-based semi-supervised strategy to align intermediate embeddings and ensure robust predictions under perturbations. To address incomplete annotations within labeled data, we design a window-based contrastive loss that fully exploits sparse and invalid depth regions. Extensive experiments across varying label ratios on the KITTI and NYU datasets demonstrate that TriConDepth consistently surpasses state-of-the-art alternatives, achieving significant improvements and showcasing the model’s robustness and effectiveness in real-world depth estimation. Code is available at https://github.com/Ketin12138/TriConDepth .