Beyond Sensing Fragility: Complementary-Pathway Distillation for Robust Depth Estimation
Abstract
Vision foundation models have achieved remarkable success in monocular depth estimation, yet they tightly couple semantic and structural cues within a single RGB observation. This coupled representation makes them fragile to real-world imaging degradations such as motion blur and exposure failure, and inherently incompatible with complementary sensing modalities that physically decouple these cues. Complementary Vision Sensors (CVS) provide a dual-pathway representation, consisting of a slow semantic pathway and a high-speed structural pathway that remain robust under adverse imaging conditions. However, directly applying RGB-foundation model to this decoupled representation leads to a fundamental modality gap, while limited CVS data prevents effective training from scratch. Here, we present a complementary-pathway distillation framework that transfers RGB foundation-model priors into the dual-pathway representation. Our approach leverages a modality-driven learning principle, where the distinct physical properties of the two pathways naturally induce semantic-structural specialization. We further introduce an asymmetric cross-modal alignment mechanism to inject stable semantic context into high-speed structural features, together with a progressive distillation strategy that stabilizes training under degraded inputs. In addition, we identify spatial-difference sensing as an effective yet underexplored modality for geometric perception. Crucially, our distillation paradigm bridges emerging brain-inspired sensing with modern computer vision. Experiments on synthetic and real-world datasets demonstrate robust, generalizable depth estimation, achieving a 757-FPS equivalent temporal resolution and maintaining a 0.26 Absolute Relative error (Abs Rel) even under severe imaging degradations.