Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

NeRF and Gaussian Splatting SLAM in the Wild

Fabian Schmidt, Markus Enzweiler, Abhinav Valada

TL;DR

This work addresses robust outdoor visual SLAM by benchmarking traditional, neural, NeRF-based, and Gaussian Splatting approaches on real outdoor data (ROVER Garden Small). It reveals that neural SLAM methods offer strong robustness under challenging lighting, albeit at high computational cost, while traditional methods excel across seasons but are sensitive to lighting changes. GO-SLAM, when paired with DROID-SLAM as an external tracker, often provides the best overall accuracy and robustness, but at substantial GPU/CPU cost; DPV-SLAM offers a lightweight real-time monocular option with competitive performance. The findings illuminate practical trade-offs for outdoor SLAM deployment and guide algorithm selection based on environmental conditions and resource budgets.

Abstract

Navigating outdoor environments with visual Simultaneous Localization and Mapping (SLAM) systems poses significant challenges due to dynamic scenes, lighting variations, and seasonal changes, requiring robust solutions. While traditional SLAM methods struggle with adaptability, deep learning-based approaches and emerging neural radiance fields as well as Gaussian Splatting-based SLAM methods, offer promising alternatives. However, these methods have primarily been evaluated in controlled indoor environments with stable conditions, leaving a gap in understanding their performance in unstructured and variable outdoor settings. This study addresses this gap by evaluating these methods in natural outdoor environments, focusing on camera tracking accuracy, robustness to environmental factors, and computational efficiency, highlighting distinct trade-offs. Extensive evaluations demonstrate that neural SLAM methods achieve superior robustness, particularly under challenging conditions such as low light, but at a high computational cost. At the same time, traditional methods perform the best across seasons but are highly sensitive to variations in lighting conditions. The code of the benchmark is publicly available at https://github.com/iis-esslingen/nerf-3dgs-benchmark.

NeRF and Gaussian Splatting SLAM in the Wild

TL;DR

This work addresses robust outdoor visual SLAM by benchmarking traditional, neural, NeRF-based, and Gaussian Splatting approaches on real outdoor data (ROVER Garden Small). It reveals that neural SLAM methods offer strong robustness under challenging lighting, albeit at high computational cost, while traditional methods excel across seasons but are sensitive to lighting changes. GO-SLAM, when paired with DROID-SLAM as an external tracker, often provides the best overall accuracy and robustness, but at substantial GPU/CPU cost; DPV-SLAM offers a lightweight real-time monocular option with competitive performance. The findings illuminate practical trade-offs for outdoor SLAM deployment and guide algorithm selection based on environmental conditions and resource budgets.

Abstract

Navigating outdoor environments with visual Simultaneous Localization and Mapping (SLAM) systems poses significant challenges due to dynamic scenes, lighting variations, and seasonal changes, requiring robust solutions. While traditional SLAM methods struggle with adaptability, deep learning-based approaches and emerging neural radiance fields as well as Gaussian Splatting-based SLAM methods, offer promising alternatives. However, these methods have primarily been evaluated in controlled indoor environments with stable conditions, leaving a gap in understanding their performance in unstructured and variable outdoor settings. This study addresses this gap by evaluating these methods in natural outdoor environments, focusing on camera tracking accuracy, robustness to environmental factors, and computational efficiency, highlighting distinct trade-offs. Extensive evaluations demonstrate that neural SLAM methods achieve superior robustness, particularly under challenging conditions such as low light, but at a high computational cost. At the same time, traditional methods perform the best across seasons but are highly sensitive to variations in lighting conditions. The code of the benchmark is publicly available at https://github.com/iis-esslingen/nerf-3dgs-benchmark.

Paper Structure

This paper contains 7 sections, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Benchmark
  4. SLAM Methods
  5. Evaluation Methodology
  6. Experiments and Results
  7. Conclusion

Figures (2)

  • Figure 1: Illustration of the ROVER dataset sequences from Garden Small location (a), captured across various environmental conditions (b), where the top row depicts the different seasons (summer, autumn, winter, spring), while the bottom row shows the varying lighting conditions (daylight, dusk, night, night with external lighting).
  • Figure 2: Qualitative trajectory comparisons for the top three Mono (GO-SLAM, DROID-SLAM, DPV-SLAM) and RGB-D (GO-SLAM, MonoGS, DROID-SLAM) SLAM methods in two scenarios, highlighting differences between the modalities. (a) Summer: Mono methods struggle with scale ambiguity and drift, while RGB-D methods also face drift problems, with DROID-SLAM and MonoGS being notably affected, whereas GO-SLAM achieves the best overall performance. (b) Night + Light: Mono methods face severe scaling problems, with GO-SLAM additionally suffering significant drift. For RGB-D, DROID-SLAM and GO-SLAM exhibit noticeable drift, while MonoGS struggles with consistent tracking and scaling.