Attention-Guided Variational Distillation for Robust Depth Estimation under Challenging Conditions

Monocular depth estimation is a low-cost visual perception technique for autonomous navigation but can suffer significant performance degradation under challenging conditions such as adverse weather and low light. Existing methods generally rely on pseudo-labels or static feature transformations, which struggle to simultaneously handle intra-domain noise and inter-domain shifts across diverse scenarios. In this work, we propose a principled and effective framework based on an enhanced information bottleneck principle and approximated via variational distillation, termed Attention-Guided Variational Distillation (AGVD).We reformulate robust depth estimation into two complementary optimization tasks that explicitly model task-related information and domain-invariant information. These two objectives are implemented through a two-stage training strategy. In the first stage, we maximize task-relevant information via Variational Self-Distillation (VSD) guided by a Dynamic Attention (DA) mechanism. In the second stage, we bridge the synthetic-to-real domain gap via Variational Cross-Distillation (VCD) guided by a Geometric Guidance (GG) mechanism. AGVD is compatible with representative self-supervised and foundation depth backbones. Extensive experiments on nuScenes, RobotCar, and DrivingStereo, together with real-world tests, demonstrate that AGVD improves depth robustness under challenging conditions. Code for this study is publicly available.