Feature-Decoupled Lightweight Monocular Depth Estimation Trained from a Single GPU
Abstract
Monocular depth estimation (MDE) predicts per-pixel depth from a single RGB image to recover the 3D structure of a scene. When trained with LiDAR supervision, MDE models can learn to predict metric-scale depth directly. However, LiDAR annotations projected onto the image plane are spatially sparse and discontinuous and may provide limited supervision near depth discontinuities and thin structures. This issue can be more challenging for lightweight networks due to limited representational capacity and repeated downsampling operations. To mitigate this issue, we propose a feature-decoupled lightweight dual-path architecture that explicitly separates structural and contextual representations during feature encoding. Based on a decouple–refine–integrate design principle, the encoder forms role-specific structural and contextual features, refines the two representations independently, and finally re-integrates contextual information under structural guidance. The proposed model can be trained from scratch on a single GPU without external pre-training while achieving competitive performance with only 2.9M parameters. Experiments on KITTI and NYU Depth v2 demonstrate competitive performance against representative lightweight backbones and dedicated MDE models. In addition, zero-shot cross-dataset evaluation from KITTI to DDAD demonstrates improved transferability across different camera configurations and outdoor driving data distributions without additional training or fine-tuning. Our code and models are publicly available at https://github.com/starsleep/FD-Depth.