SceneCoder-SLAM: End-to-End Semantic Visual SLAM via Differentiable Bundle Adjustment and Learnable Implicit Representations
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Traditional Simultaneous Localization and Mapping (SLAM) systems typically follow modular pipelines that treat semantic understanding, pose estimation, and map construction as separate stages. We propose SceneCoder-SLAM, a unified end-to-end learnable framework that tightly integrates neural implicit scene representations with a Differentiable Bundle Adjustment (DBA) layer. Unlike existing methods that separate feature extraction from optimization, our framework enables joint optimization of semantic and geometric information through gradient-based learning. Key contributions include: (1) a multi-task auto-encoder with orthogonal feature decoupling to isolate geometric and semantic latent codes; (2) a fully differentiable bundle adjustment layer that leverages implicit neural fields for continuous scene modeling and propagates residuals to the frontend; and (3) a self-supervised training strategy requiring only monocular RGB sequences. Extensive experiments on the TUM RGB-D, ScanNet, and Replica datasets demonstrate that SceneCoder-SLAM reduces Absolute Trajectory Error (ATE) by 31.5% compared to ORB-SLAM3 and by 18.7% compared to the learning-based baseline DROIDSLAM. Furthermore, we provide new analyses demonstrating robustness to low-resolution inputs and justifying the computational efficiency of our decoupling mechanism. The results indicate a shift toward holistic, data-driven SLAM systems capable of robust performance in dynamic environments.
Fields of Study & Subjects
Robotics and Sensor-Based Localization
3D Surveying and Cultural Heritage
Advanced Vision and Imaging
Publication details
- Journal / source
- International Conference on Electrical Automation and Artificial Intelligence (ICEAAI)
- Pages
- 901–904
- Publication date
- 16 January 2026
- Language
- EN