Dynamic-Parameterized Reconstruction Model for Resource-Aware Spatial Intelligence.

Abstract

Spatial intelligence in autonomous driving requires object-level 3D geometry, yet existing monocular mesh reconstruction methods usually operate with a fixed inference path and a single mesh parameterization, which limits their flexibility under heterogeneous resource constraints. To address this issue, we propose DyPRSI, a dynamic-parameterized framework for monocular vehicle 3D reconstruction that provides multiple predefined accuracy-latency operating points within a single model. DyPRSI inserts two early exits into a shared Res2Net-BiFPN trunk and associates each exit with an exit-specific mesh specification, forming a coarse-to-fine reconstruction hierarchy across network depth. To better match the efficiency requirements of shallow branches, DyPRSI adopts lightweight coordinate-classification keypoint decoding for EE1 and EE2, while retaining a heatmap-regression keypoint head in the Main branch to preserve the upper bound of reconstruction accuracy. Experiments on ApolloCar3D show that DyPRSI-Main achieves competitive reconstruction performance, whereas EE1 and EE2 substantially reduce end-to-end inference latency and provide useful alternatives under different resource requirements. Ablation studies further show that the speedup mainly comes from the lightweight branch-specific keypoint heads, while the exit-specific mesh settings help organize stable coarse-to-fine reconstruction behavior across branches. These results indicate that DyPRSI is a practical monocular vehicle reconstruction framework for resource-aware spatial intelligence.