Huixin Zhu, Junyang Zhao, Zhili Zhang, Weisheng Dong, Mingtao Feng, Han Yu. MDViT-SLAM: Multi-scale Dense Vision Transformer Enhanced 3D Gaussian Splatting for Monocular VSLAMJ. Machine Intelligence Research, 2026, 23(6): 1283-1301. DOI: 10.1007/s11633-025-1611-z
Citation: Huixin Zhu, Junyang Zhao, Zhili Zhang, Weisheng Dong, Mingtao Feng, Han Yu. MDViT-SLAM: Multi-scale Dense Vision Transformer Enhanced 3D Gaussian Splatting for Monocular VSLAMJ. Machine Intelligence Research, 2026, 23(6): 1283-1301. DOI: 10.1007/s11633-025-1611-z

MDViT-SLAM: Multi-scale Dense Vision Transformer Enhanced 3D Gaussian Splatting for Monocular VSLAM

  • Visual simultaneous localization and mapping (VSLAM) enables autonomous navigation through precise pose estimation and dense 3D reconstruction. However, existing monocular VSLAM systems suffer from detailed structural loss, limited trajectory accuracy, and poor reconstruction accuracy. This extremely limits their capability to autonomous navigation. To address these challenges, we propose multi-scale dense vision transformer (MDViT)-SLAM, the first dense monocular framework which employs a hierarchical cross-attention mechanism to efficiently estimate the depth of the monocular data and incorporates 3D Gaussians for efficient scene representation. To mitigate depth inaccuracy and scale drift in monocular input, we propose a multi-scale vision transformer (MViT) that integrates pre-trained ViT encoders and cross-scale fusion modules, concurrently capturing global contexts and local geometries. For enhanced scene reconstruction fidelity, we introduce learnable anisotropic Gaussians into SLAM to serve as foundational representation units, enabling adaptive geometric parameterization and real-time rendering. To overcome sparse reconstruction artifacts in continuous visual stream processing, we develop an adaptive Gaussians densification strategy addressing reconstruction efficiency via systematic redundancy pruning, achieving centimeter-level precision. Evaluations on the TUM-RGB-D fr2 dataset show that MDViT-SLAM attains 1.41 cm ATE, representing a 71% improvement over baseline models, with a color optimization rendering performance of 1078.43 FPS. Extensive experiments demonstrate the framework′s superior scene representation capabilities and computational efficiency, validating its applicability in autonomous driving and high-definition mapping.
  • loading

Catalog

    /

    DownLoad:  Full-Size Img  PowerPoint
    Return
    Return