IG-SLAM: Instant Gaussian SLAM (2408.01126v2)

Abstract: 3D Gaussian Splatting has recently shown promising results as an alternative scene representation in SLAM systems to neural implicit representations. However, current methods either lack dense depth maps to supervise the mapping process or detailed training designs that consider the scale of the environment. To address these drawbacks, we present IG-SLAM, a dense RGB-only SLAM system that employs robust Dense-SLAM methods for tracking and combines them with Gaussian Splatting. A 3D map of the environment is constructed using accurate pose and dense depth provided by tracking. Additionally, we utilize depth uncertainty in map optimization to improve 3D reconstruction. Our decay strategy in map optimization enhances convergence and allows the system to run at 10 fps in a single process. We demonstrate competitive performance with state-of-the-art RGB-only SLAM systems while achieving faster operation speeds. We present our experiments on the Replica, TUM-RGBD, ScanNet, and EuRoC datasets. The system achieves photo-realistic 3D reconstruction in large-scale sequences, particularly in the EuRoC dataset.

• The paper introduces IG-SLAM, a dense RGB SLAM system that combines DROID-SLAM tracking with Gaussian Splatting mapping for real-time 3D reconstruction.
• It leverages depth uncertainty to optimize Gaussian parameters, achieving photorealistic outputs with high PSNR and SSIM metrics.
• The system operates at nearly 10 fps using a sliding window strategy, outperforming earlier methods in both speed and visual quality.

IG-SLAM: Instant Gaussian SLAM

The paper "IG-SLAM: Instant Gaussian SLAM" by F. Aykut Sarıkamış and A. Aydın Alatan presents a novel approach to Simultaneous Localization and Mapping (SLAM) leveraging dense RGB data and Gaussian Splatting. Herein, I delineate the notable aspects and broader implications of this work.

Overview

IG-SLAM is a dense RGB-only SLAM system that combines robust dense SLAM methods for tracking with Gaussian Splatting for mapping. The salient features that delineate this system from its contemporaries are:

1. Robust Pose Estimation: It employs highly accurate pose estimation using DROID-SLAM as its tracking module.
2. Depth Uncertainty Utilization: The system incorporates depth uncertainty to enhance 3D reconstruction, making mapping robust to noise.
3. Efficient Mapping Algorithm: Designed to operate at a high frame rate of 10 fps in a single process, the mapping algorithm utilizes a decay strategy and a sliding window optimization approach.

Contributions

The authors outline several key contributions which can be summarized as follows:

1. IG-SLAM Architecture: A robust dense RGB SLAM system that combines dense depth maps and Gaussian Splatting to deliver high-frame-rate performance.
2. 3D Reconstruction Algorithm: An innovative approach that accounts for depth uncertainty, ensuring robustness against noise in the 3D reconstruction.
3. Training Procedure: A detailed training methodology aiming at efficient use of dense depth supervision in the mapping process.

Methodology

The methodology is broken down into two primary components: tracking and mapping.

Tracking

The tracking approach leans on DROID-SLAM, which estimates camera poses and dense depth maps. The process involves constructing a frame graph based on co-visibility, followed by dense bundle adjustment (DBA) to minimize reprojection error. The system performs global bundle adjustment periodically to mitigate drift.

Mapping

The mapping component is centered on Gaussian Splatting. Gaussians are initialized from the keyframe's depth map, utilizing a covariance mask to account for depth uncertainty. A differentiable rendering pipeline is then employed, where the loss function is a combination of weighted depth loss and color loss. The positions, orients, scales, opacities, and colors of the Gaussians are iteratively optimized.

Experiments and Results

The authors conducted comprehensive experiments across various datasets, including Replica, TUM-RGBD, ScanNet, and EuRoC, demonstrating IG-SLAM's competitive performance in terms of visual quality and real-time operation speed.

1. Rendering and Reconstruction Accuracy: IG-SLAM shows competitive performance, achieving photorealistic 3D reconstruction with high PSNR and SSIM values.
2. Runtime Performance: The system operates at a significantly higher frame rate (9.94 fps) compared to other methods like Splat-SLAM (1.24 fps).

Implications

Practically, IG-SLAM facilitates real-time, high-fidelity 3D scene reconstruction using only RGB data, which is particularly valuable for applications in augmented reality, robotics, and large-scale environment mapping. Theoretically, it combines the strengths of dense SLAM algorithms with advanced 3D representation techniques, paving the way for more integrated and efficient SLAM systems.

Limitations and Future Work

While the system demonstrates robust performance, one limitation mentioned is the blurry edges arising from upsampled dense depth maps used in tracking. Future iterations could explore more sophisticated depth map refinement techniques to address this issue. Additionally, incorporating more advanced depth estimation and uncertainty quantification methods could further enhance IG-SLAM's robustness and accuracy.

Conclusion

IG-SLAM represents a significant advancement in the SLAM landscape by integrating robust dense tracking methods with Gaussian Splatting for high-fidelity, real-time 3D reconstruction using RGB data. Its novel use of depth uncertainty and efficient mapping algorithm positions it as a promising approach for various real-world applications, setting the stage for future developments in dense SLAM systems.