Thermal Depth Estimation Using Unified Multi-Scale Features and Propagation-Based Refinement

Thermal monocular depth estimation can provide more robust depth predictions than RGB-based methods under nighttime and adverse weather conditions. However, when trained with projected LiDAR supervision, depth models often retain structural errors in sky regions, long-range areas, and object boundaries because LiDAR measurements are sparse or missing in such regions. To address this limitation, we propose a thermal monocular depth estimation framework that incorporates propagation-based refinement. To make this refinement applicable across different base models, we further design a multi-scale feature adapter that converts heterogeneous multi-scale features with different spatial resolutions and channel dimensions into a unified representation. As a result, the same refinement architecture can be used across different base models without model-specific refiner redesign. On the multispectral stereo (MS2) dataset, the proposed method improves both BTS (big-to-small) and NeWCRFs (neural window fully connected CRFs), reducing the meter-based error metrics SqRel from 0.380 to 0.369 and RMSE from 3.163 to 3.126 for BTS, and reducing SqRel from 0.331 to 0.328 and RMSE from 2.937 to 2.924 for NeWCRFs. Qualitative results further show that the proposed method alleviates mixed-depth artifacts and abnormal depth patterns in regions lacking reliable depth supervision.