Teacher–student-paradigm-based 3D range-gated imaging method: exploiting depth cues from pre-trained large RGB depth estimation models

3D range-gated imaging (3DRGI) has great potential for long-range detection in adverse weather and nighttime low-light conditions, though existing methods are limited by the inherent flaws of gated images, such as a lack of color and susceptibility to specular noise, which degrades depth accuracy. Concurrently, large depth estimation models with millions of parameters, trained on large-scale RGB images have demonstrated powerful generalization for depth estimation. To bridge this gap, we propose a teacher–student paradigm that transfers robust relative depth knowledge from a pre-trained large RGB depth model (teacher) to a 3DRGI model (student). We first fine-tune the teacher model and distill its scene-level understanding using a relative-to-metric depth conversion strategy. Then, we introduce relative depth and edge consistency losses to guide our student model’s training, ensuring that its predictions remain aligned with the teacher’s strong priors. To further enhance depth quality and mitigate overfitting to sparse LiDAR features, we also propose a center-jitter cropping strategy. Extensive experiments and ablation studies validate our approach, demonstrating that it achieves state-of-the-art performance with comprehensive improvements across accuracy, visual quality, and computational efficiency. This cross-modal guidance paradigm highlights a promising direction for enhancing perception systems by leveraging large-scale models across different sensor domains.