Self-Supervised Geometry-Guided Initialization for Robust Monocular Visual Odometry

Takayuki Kanai, Igor Vasiljevic, Vitor Guizilini, Kazuhiro Shintani

TL;DR

This paper introduces SG-Init, a novel initialization strategy for monocular visual odometry that leverages self-supervised depth and pose estimation priors. By providing a more robust initial guess to dense bundle adjustment, SG-Init significantly improves accuracy and robustness in challenging scenarios like large motions and dynamic objects, without requiring modifications to the underlying SLAM backbone.

What Problem Does This Paper Address?

Monocular visual odometry (VO) is crucial for autonomous systems, but traditional feature-based methods struggle with poor lighting, lack of texture, and rapid movements. While recent learning-based dense SLAM methods, particularly those using dense bundle adjustment (DBA), have shown promise, they still falter in scenarios with large ego-motion or dynamic objects. This is often due to the optimization process converging to a suboptimal solution early on, especially when the initial guess for depth and pose is inaccurate. Existing methods often resort to domain-specific fine-tuning of the SLAM backbone, which sacrifices generality. This paper aims to address these limitations by improving the initialization of the DBA process, making it more robust and generalizable.

Methods

The core of the proposed method, SG-Init, lies in its novel initialization strategy for dense bundle adjustment. Instead of relying on potentially noisy or inaccurate initial guesses, SG-Init leverages self-supervised geometric priors. Specifically:

  • Self-Supervised Depth and Ego-Motion Learning: The authors utilize pre-trained "DepthNet" and "PoseNet" models trained using self-supervision. These models predict depth and relative camera poses from video sequences.
  • Zero-Shot Depth Guidance: To mitigate the inherent ambiguities and potential inaccuracies in self-supervised depth estimation (e.g., infinite depth predictions, edge bleeding), SG-Init incorporates a large-scale, pre-trained zero-shot monocular depth estimator. This provides a more reliable and scale-aware depth prior.
  • Geometric Initialization: The predicted depth from the zero-shot model and the relative poses from the self-supervised PoseNet are combined to form an initial guess for the dense bundle adjustment (DBA) layer. This initial guess is then used to refine the camera poses and depth maps.
  • Integration with DROID-SLAM: The proposed SG-Init is designed to be plug-and-play, demonstrated by its integration with the DROID-SLAM framework. The key idea is to improve the initialization of the DBA step within DROID-SLAM without altering its core architecture or requiring further training.

Datasets

The paper uses the following datasets:

  • DDAD (Dense Depth for Autonomous Driving): Used for training and evaluation. It features synchronized RGB cameras and dense LiDAR point clouds from diverse driving scenes.
  • KITTI: A standard benchmark used for both training (monocular RGB images) and evaluation of visual odometry.

Key Insights and Ideas

  • Initialization is Key: The paper strongly argues that the performance bottleneck in learning-based dense SLAM often lies not in the core optimization algorithm itself, but in the quality of its initialization, especially in challenging scenarios.
  • Zero-Shot Priors are Powerful: Leveraging pre-trained, zero-shot depth estimators significantly improves robustness, even when the underlying SLAM backbone is not fine-tuned. This suggests that general geometric understanding from large-scale data is highly beneficial.
  • Depth and Pose Initialization Synergy: The authors demonstrate that initializing both depth and pose is crucial for optimal performance, rather than relying solely on improved depth estimation. This highlights the interconnectedness of these geometric cues in the optimization process.
  • Generality Preserved: By focusing on initialization rather than backbone fine-tuning, SG-Init maintains the generality of the SLAM system, allowing it to perform well across different environments without specialized training.

Key Results

  • Significant Accuracy Improvements: SG-Init achieves substantial improvements in Absolute Trajectory Error (ATE) on both the challenging DDAD benchmark (e.g., reducing ATE from 6.007 to 0.451 with DROID-SLAM) and the KITTI benchmark.
  • Robustness in Challenging Scenarios: The method shows marked improvements in scenarios with large ego-motion, dynamic objects, and textureless regions, as visualized in Figure 5 and quantified in Table I.
  • Computational Efficiency: SG-Init is approximately ten times faster than other state-of-the-art methods while maintaining competitive accuracy and low GPU RAM usage (Table IV).
  • Competitive with SOTA: The proposed method achieves competitive or superior trajectory error results compared to established methods, even in a monocular setup.

Limitations and Future Work

  • Memory Footprint: As SG-Init relies on dense bundle adjustment, it inherently requires more memory than traditional sparse methods.
  • Reliance on Learned Priors: The performance of SG-Init is dependent on the quality of the self-supervised depth and ego-motion models.
  • Indoor Environments: The current work focuses on outdoor driving scenarios. The authors suggest that applying SG-Init to indoor environments, which often have varying depth ranges and significant rotational motions, might degrade performance and is an interesting area for future research.
  • Generalization to Diverse Scenarios: Further research is needed to ensure optimal accuracy in general settings (e.g., indoors) while maintaining the capability demonstrated in driving scenarios.

Summary updated

Abstract

Monocular visual odometry is a key technology in various autonomous systems. Traditional feature-based methods suffer from failures due to poor lighting, insufficient texture, and large motions. In contrast, recent learning-based dense SLAM methods exploit iterative dense bundle adjustment to address such failure cases, and achieve robust and accurate localization in a wide variety of real environments, without depending on domain-specific supervision. However, despite its potential, the methods still struggle with scenarios involving large motion and object dynamics. In this study, we diagnose key weaknesses in a popular learning-based dense SLAM model (DROID-SLAM) by analyzing major failure cases on outdoor benchmarks and exposing various shortcomings of its optimization process. We then propose the use of self-supervised priors leveraging a frozen large-scale pre-trained monocular depth estimator to initialize the dense bundle adjustment process, leading to robust visual odometry without the need to fine-tune the SLAM backbone. Despite its simplicity, the proposed method demonstrates significant improvements on KITTI odometry, as well as the challenging DDAD benchmark.
BibTeX
@inproceedings{kanai2025self,
  author = {Takayuki Kanai and Igor Vasiljevic and Vitor Guizilini and Kazuhiro Shintani},
  title = {{Self-Supervised Geometry-Guided Initialization for Robust Monocular Visual Odometry}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2025},
  doi = {10.1109/iros60139.2025.11247022},
}
Figure