Noise-Aware 3D Scene Reconstruction via Diffusion-Auxiliary Implicit SLAM

Neural surface reconstruction has recently gained increasing attention due to its promising results in high-fidelity scene modeling and rendering. Since implicit neural rendering has archived superior performance in scene reconstruction, existing visual SLAM approaches still face significant challenges in real-time reconstruction due to uneven exposure, motion blur, and sensor-induced noise, which are collectively referred to as noisy imaging conditions. In this paper, we propose an adaptive implicit SLAM framework robust to noisy environments, which integrates an auxiliary diffusion model that supports both an image restoration network and a noise-aware sampling strategy, enhancing the system’s robustness to degraded inputs and improving both tracking and mapping performance. Specifically, our system adopts a multilayer perceptron (MLP) combined with auxiliary diffusion modules to simultaneously perform pose estimation and indoor scene reconstruction in real time. Additionally, we incorporate a truncated signed distance function to enhance reconstruction accuracy and robustness. Experiments on several challenging benchmark datasets demonstrate that our method outperforms several methods in noisy conditions.