Monocular MapTR: A Depth-Aware Approach to HD-Map Extraction from a Single Camera

Recent progress in vectorized HD-map extraction has been driven by end-to-end transformer architectures such as MapTR and MapTR-v2, which typically assume access to synchronized multi-view cameras. In this work, we extend MapTR to the monocular setting, enabling the detection of map elements from a single front-facing camera. To achieve this, we introduce three modifications to the standard pipeline: (i) a camera-aware BEV lifting stage conditioned on learnable view tokens; (ii) the integration of monocular depth priors obtained from MiDaS, discretized into an uncertainty-aware cost volume that guides BEV features; and (iii) a training strategy tailored to monocular coverage, incorporating temporal and geometric augmentations together with a short-cycle schedule that ensures stable convergence without the need for warmup. The decoder, set-prediction losses, and vectorized targets remain consistent with MapTR for simplicity. Evaluated on the nuScenes validation set, the proposed monocular MapTR achieves competitive performance across lane dividers, pedestrian crossings, and boundaries (measured by Chamfer-based AP). It substantially reduces the gap to multi-camera models and clearly outperforms a naïve single-camera baseline. Ablation studies further reveal that the depth priors derived from MiDaS and the introduction of camera tokens account for most of the improvements, while refinements to the training schedule enhance both stability and sample efficiency. Taken together, these results demonstrate a practical and cost-effective path to HD-map extraction in scenarios where only one camera is available.