Abstract
Pose Graph Optimization (PGO) is a critical component in 3D reconstruction and SLAM, traditionally solved via iterative optimization. This paper introduces PoGO-Net, a Graph Neural Network (GNN) framework that reformulates PGO as a message-passing problem. By incorporating a differentiable "edge-dropping" mechanism, PoGO-Net effectively filters outlier constraints that typically degrade performance. Experimental results demonstrate that PoGO-Net achieves state-of-the-art accuracy across diverse benchmarks while operating up to 400x faster than traditional CPU-based solvers.
Introduction
In computer vision tasks such as Structure-from-Motion (SfM), PGO reconciles noisy, pairwise relative camera poses into a globally consistent coordinate frame. Traditional solvers face significant challenges: high computational complexity (often cubic), extreme sensitivity to outlier measurements from poor feature matching, and limited generalization across different environments. PoGO-Net addresses these by leveraging the parallel processing power of GNNs and an integrated denoising strategy.
Methodology
PoGO-Net represents the optimization problem as a graph where nodes are absolute camera orientations and edges are relative rotation constraints.
Differentiable Edge-Dropping
The core innovation is the De-noise Layer. Unlike standard GNNs that aggregate all neighbor information, PoGO-Net learns to predict the reliability of each edge. Using the Gumbel-Softmax trick, the network makes discrete "keep or drop" decisions differentiable. This allows the model to learn outlier rejection end-to-end during training based on local graph consistency.
Iterative Optimization
The network iteratively updates node features through message aggregation. It is initialized using a spanning tree starting from the most connected camera node to provide a stable starting point. Training is guided by a joint loss function that enforces both global geometric consistency and alignment with ground-truth poses.
Experimental Results
The framework was evaluated on benchmarks including 7Scenes, Cambridge, ScanNet, and Photo Tourism.
Accuracy and Robustness
PoGO-Net consistently outperformed both traditional robust solvers (e.g., IRLS) and previous learning-based methods. On the 7Scenes dataset, it reduced the average angular error to 4.93°, a significant improvement over the 6.33° achieved by prior state-of-the-art learning methods.
Computational Efficiency
The use of GNNs allows for massive parallelism on GPUs. In large-scale Photo Tourism datasets, PoGO-Net demonstrated speeds up to 400x faster than conventional CPU solvers and 2x faster than competing GNN-based approaches, making it suitable for real-time applications.
Conclusion
PoGO-Net represents a significant step forward in robust, real-time pose graph optimization. By integrating denoising directly into the GNN architecture, it overcomes the sensitivity to outliers that plagues traditional methods. Future work includes extending the framework to full $SE(3)$ regression (handling both rotation and translation) and integration into end-to-end SLAM systems.
Summary updated