9 June 2026 in SSRN Electronic Journal
Monocular 3D vehicle shape and pose estimation is of great significance for scene understanding in autonomous driv-ing. However, it encounters challenges owing to the lack of depth information, complex interactions between objects, and efficient modeling requirements. Unlike existing monocular methods that often rely on either coarse PCA-based shape priors or computationally expensive attention mechanisms, this paper proposes PFHI, an end-to-end frame-work designed to alleviate the accuracy–efficiency trade-off in joint monocular vehicle shape–pose reconstruction. First, we design a prior-factor-driven attention to effectively aggregate prior information, thereby improving shape reconstruction accuracy while notably reducing computational cost. Second, we introduce a hierarchical interactive attention module that decouples object–object and object–scene interactions, enhancing the structural representation for pose estimation. In addition, Dynamic Tanh (DyT) is incorporated as a lightweight normalization alternative to improve convergence stability and inference efficiency without relying on statistic-dependent normalization opera-tions. Experimental results on the ApolloCar3D dataset show that, compared to the existing state-of-the-art model, the proposed method improves the A3DP-Abs and A3DP-Rel indicators by 1.87% and 6.26%, respectively, while reducing inference time by 41.46%.
27 May 2026 in SSRN Electronic Journal
Monocular 3D vehicle shape and pose estimation is of great significance for scene understanding in autonomous driving. However, it encounters challenges owing to the lack of depth information, complex interactions between objects, and efficient modeling requirements. In this paper, we propose an end-to-end framework that integrates prior factors and hierarchical interaction to jointly model the 3D shape and 6 DOF pose of a vehicle. First, we design a prior-factor-driven attention to effectively aggregate prior information, thereby improving shape reconstruction accuracy while notably reducing computational cost. Second, we introduce a hierarchical interactive attention module that decouples object–object and object–scene interactions, enhancing the structural representation for pose estimation. We replace the traditional normalization operation with Dynamic Tanh (DyT) for the first time, further improving the model’s convergence stability and inference efficiency. Experimental results on the ApolloCar3D dataset show that, compared to the existing state-of-the-art model, the proposed method improves the AD3P-Abs and AD3P-Rel indicators by 1.87% and 6.26%, respectively, while reducing inference time by 41.46%.