NeuV-SLAM: Fast Neural Multiresolution Voxel Optimization for RGBD Dense SLAM (2402.02020v1)

Abstract: We introduce NeuV-SLAM, a novel dense simultaneous localization and mapping pipeline based on neural multiresolution voxels, characterized by ultra-fast convergence and incremental expansion capabilities. This pipeline utilizes RGBD images as input to construct multiresolution neural voxels, achieving rapid convergence while maintaining robust incremental scene reconstruction and camera tracking. Central to our methodology is to propose a novel implicit representation, termed VDF that combines the implementation of neural signed distance field (SDF) voxels with an SDF activation strategy. This approach entails the direct optimization of color features and SDF values anchored within the voxels, substantially enhancing the rate of scene convergence. To ensure the acquisition of clear edge delineation, SDF activation is designed, which maintains exemplary scene representation fidelity even under constraints of voxel resolution. Furthermore, in pursuit of advancing rapid incremental expansion with low computational overhead, we developed hashMV, a novel hash-based multiresolution voxel management structure. This architecture is complemented by a strategically designed voxel generation technique that synergizes with a two-dimensional scene prior. Our empirical evaluations, conducted on the Replica and ScanNet Datasets, substantiate NeuV-SLAM's exceptional efficacy in terms of convergence speed, tracking accuracy, scene reconstruction, and rendering quality.

• The paper introduces NeuV-SLAM, a dense SLAM framework that uses neural multiresolution voxels for dynamic scene expansion and rapid convergence.
• It leverages a novel hashMV voxel management and an implicit VDF representation to capture fine scene details and enhance tracking accuracy.
• Experimental results on the Replica and ScanNet datasets demonstrate lower RMSE and superior mapping quality compared to traditional SLAM methods.

Introduction

Simultaneous Localization and Mapping (SLAM) is a well-established problem in computational perception with wide-ranging applications, from autonomous vehicle navigation to augmented reality. Dense SLAM, in particular, aims to build richly detailed maps of the environment while tracking the camera's position. Traditional SLAM systems often struggle to capture intricate details, particularly regarding color and texture consistency, due to their reliance on explicit spatial representations. Implicit neural representations have emerged as a powerful tool for rendering photorealistic scenes, yet current NeRF-based SLAM methods face challenges in rapid and efficient incremental scene expansion and convergence.

NeuV-SLAM Methodology

The paper introduces NeuV-SLAM, a novel dense SLAM framework based on neural multiresolution voxels that can efficiently expand and learn scenes from sequential RGBD frames. Key to the system is a new hash-based multiresolution voxel management called hashMV, which enables dynamic scene expansion. Simultaneously, a new implicit representation called VDF is proposed to convert SDF values and color features within voxels, which significantly enhances the capacity to capture finer scene details and improve convergence efficiency.

Evaluation and Results

The paper's empirical evaluations leveraging the Replica and ScanNet Datasets demonstrate the system's strengths. The results from these datasets verify that NeuV-SLAM showcases exceptional convergence speed and superior tracking accuracy compared to other contemporary systems. It also displays an advanced capacity for quality scene reconstruction and rendering. Specifically, its ability to derive detailed maps and maintain accurate camera tracking in large and complex environments was highlighted by strong numerical results, with lower root-mean-square error (RMSE) in localization and higher accuracy and completion rates in reconstruction compared to other leading methods.

Ablation Studies and Performance

Ablation studies further confirm the importance of both the SDF activation strategy and the use of multiresolution voxels. Removing SDF activation leads to a loss of detail and a marked decrease in performance metrics. Similarly, employing single-resolution voxels results in a degradation of both tracking and mapping accuracy, as well as increased memory usage. On the flip side, different threshold values used for edge detection in voxel generation also impact system performance, with lower thresholds leading to better results, possibly due to creating more dense voxels that positively influence the SLAM process.

Conclusion

This research presents a significant step forward in dense SLAM technology. NeuV-SLAM introduces advanced strategies for voxel management and scene representation, setting new benchmarks in convergence speed and scene reconstruction. These innovations prove particularly effective in scenarios demanding high fidelity in spatial and color details. Concluding, NeuV-SLAM paves the way for more capable and scalable SLAM applications, a promising development for future autonomous systems and virtual interaction technologies.