Pose Graph Optimization for Unsupervised Monocular Visual Odometry

Yang Li, Yoshitaka Ushiku, Tatsuya Harada

Abstract

Unsupervised Learning based monocular visual odometry (VO) has lately drawn significant attention for its potential in label-free leaning ability and robustness to camera parameters and environmental variations. However, partially due to the lack of drift correction technique, these methods are still by far less accurate than geometric approaches for large-scale odometry estimation. In this paper, we propose to leverage graph optimization and loop closure detection to overcome limitations of unsupervised learning based monocular visual odometry. To this end, we propose a hybrid VO system which combines an unsupervised monocular VO called NeuralBundler with a pose graph optimization back-end. NeuralBundler is a neural network architecture that uses temporal and spatial photometric loss as main supervision and generates a windowed pose graph consists of multi-view 6DoF constraints. We propose a novel pose cycle consistency loss to relieve the tensions in the windowed pose graph, leading to improved performance and robustness. In the back-end, a global pose graph is built from local and loop 6DoF constraints estimated by NeuralBundler and is optimized over SE(3). Empirical evaluation on the KITTI odometry dataset demonstrates that 1) NeuralBundler achieves state-of-the-art performance on unsupervised monocular VO estimation, and 2) our whole approach can achieve efficient loop closing and show favorable overall translational accuracy compared to established monocular SLAM systems.

Introduction

Visual Odometry (VO) is a fundamental task in robotics and computer vision, enabling autonomous systems to estimate their trajectory from a sequence of images. While classical geometric methods have achieved high precision, they often require manual tuning and can be sensitive to environmental factors like lighting and texture. Unsupervised deep learning approaches offer a robust, data-driven alternative that does not require ground-truth labels. However, these learning-based methods typically suffer from significant cumulative drift over long sequences because they lack global consistency mechanisms like loop closure.

This paper presents a hybrid system that integrates the robustness of unsupervised learning with the global consistency of classical pose graph optimization. By combining a novel neural front-end, NeuralBundler, with a graph-based back-end, the proposed method achieves state-of-the-art accuracy while effectively mitigating long-term drift.

Methodology

NeuralBundler Front-End

The front-end of the system, NeuralBundler, processes a sliding window of $N$ frames to estimate both depth and relative motion. Unlike standard methods that only consider consecutive frames, NeuralBundler predicts 6-Degrees-of-Freedom (6DoF) constraints between all pairs of frames within the window.

A key contribution is the Pose Cycle Consistency Loss. This loss regularizes the network by enforcing that the product of relative transformations along any closed loop within the window (e.g., Frame A $\rightarrow$ B $\rightarrow$ C $\rightarrow$ A) equals the identity matrix. This constraint forces the network to produce internally consistent motion estimates, significantly improving local odometry performance.

Global Pose Graph Back-End

To address long-term drift, the system incorporates a global pose graph optimization back-end. It utilizes a loop closure detection module (DBoW2) to identify previously visited locations. When a loop is detected, the system adds a 6DoF constraint to the global graph. The entire trajectory is then optimized over $SE(3)$ using non-linear least squares, ensuring that the local estimates from NeuralBundler align with the global loop closure constraints.

System architecture overview showing the integration of the NeuralBundler front-end with the Pose Graph Optimization back-end.

Experimental Results

The system was evaluated on the KITTI odometry dataset. During training, stereo pairs were used to establish absolute metric scale, though the system operates in a purely monocular mode during inference.

Performance and Drift Correction

NeuralBundler achieved state-of-the-art results among unsupervised monocular VO methods, outperforming established baselines like SfMLearner and UnDeepVO. The integration of the pose graph back-end proved highly effective; for instance, on KITTI sequence 00, loop closure optimization reduced translational error from 4.33% to 3.24%. Furthermore, the system demonstrated competitive performance against the feature-based ORB-SLAM, achieving lower Absolute Trajectory Error (RMSE) on 8 out of 11 sequences.

Comparison of the proposed optimized trajectory (blue) against the ground truth (red dashed) on the KITTI dataset, demonstrating high accuracy and minimal drift.

Conclusion

This work demonstrates the power of hybrid systems in visual odometry. By leveraging unsupervised learning for robust local feature extraction and motion estimation, and classical graph optimization for global consistency, the proposed approach overcomes the primary limitation of learning-based VO: long-term drift. The results on the KITTI dataset highlight the potential for such hybrid systems to surpass traditional geometric SLAM in diverse real-world environments.

Summary updated

Abstract

Unsupervised Learning based monocular visual odometry (VO) has lately drawn significant attention for its potential in label-free leaning ability and robustness to camera parameters and environmental variations. However, partially due to the lack of drift correction technique, these methods are still by far less accurate than geometric approaches for large-scale odometry estimation. In this paper, we propose to leverage graph optimization and loop closure detection to overcome limitations of unsupervised learning based monocular visual odometry. To this end, we propose a hybrid VO system which combines an unsupervised monocular VO called NeuralBundler with a pose graph optimization back-end. NeuralBundler is a neural network architecture that uses temporal and spatial photometric loss as main supervision and generates a windowed pose graph consists of multi-view 6DoF constraints. We propose a novel pose cycle consistency loss to relieve the tensions in the windowed pose graph, leading to improved performance and robustness. In the back-end, a global pose graph is built from local and loop 6DoF constraints estimated by NeuralBundler and is optimized over SE(3). Empirical evaluation on the KITTI odometry dataset demonstrates that 1) NeuralBundler achieves state-of-the-art performance on unsupervised monocular VO estimation, and 2) our whole approach can achieve efficient loop closing and show favorable overall translational accuracy compared to established monocular SLAM systems.
BibTeX
@inproceedings{li2019pose,
  author = {Yang Li and Yoshitaka Ushiku and Tatsuya Harada},
  title = {{Pose Graph Optimization for Unsupervised Monocular Visual Odometry}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2019},
  url = {https://openreview.net/forum?id=BVrShuRL1SrA},
}
Figure