Knowledge Distillation with Context-Aware Refinement for Self-Supervised Depth Estimation in Transparent Environments

In recent years, self-supervised depth estimation methods have been utilized in various applications such as virtual reality, robotics, and autonomous driving. Although depth estimation has been extensively studied, existing methods still struggle to accurately predict the depth of transparent objects. Autonomous driving is often conducted in outdoor environments where transparent objects such as glass are frequently encountered. Inaccurate depth estimation for such objects can severely impact downstream tasks in autonomous driving, such as obstacle avoidance, localization and path planning, ultimately leading to unsafe navigation decisions. We propose a simple and efficient method using foundation models to generate segmentation masks that capture contextual cues around transparent objects like glass doors. The proposed context-aware refinement module leverages depth similarity in adjacent wall regions, and this depth cue is further used in a self-distillation framework to enhance depth estimation in transparent regions. Our proposed approach is evaluated on a real-world dataset that includes various objects, including glass doors. By integrating our method with an existing model, the absolute relative error in the transparent region is reduced by 23.5%, improving from 0.145 to 0.111. We also release our dataset, which was collected under guided navigation scenarios between buildings, with each sequence containing various views of glass doors. The code and our dataset are available at https://github.com/Jmyeong/CRM.