- The paper proposes using semidefinite relaxations with a truncated least squares cost function for robust multiview triangulation against noise and outliers.
- Two novel formulations are introduced: one based on epipolar constraints for efficiency under moderate noise, and another on fractional reprojection constraints for robustness against severe outliers.
- Empirical results show these methods achieve provably optimal and robust triangulations, maintaining tightness of relaxations even under high noise and outlier conditions.
Semidefinite Relaxations for Robust Multiview Triangulation
Multiview triangulation, an integral process in computer vision, involves estimating the precise 3D location of a point from multiple 2D observations across different views. This estimation is crucial for applications in structure-from-motion, augmented reality, and photogrammetry. Given the intrinsic inaccuracies in measurement due to noise and outliers, obtaining reliable triangulations is challenging. This paper presents a novel approach by incorporating semidefinite relaxations to achieve certifiably optimal robust triangulations.
Problem Definition and Challenges
Traditionally, the multiview triangulation problem is modeled as a nonconvex optimization task to minimize reprojection errors. This nontrivial problem is often corrupted by noise and outliers stemming from image feature matching, leading to suboptimal local solutions when tackled using standard methods. Past efforts have seen the application of semidefinite relaxations to non-robust triangulations, yet they falter under high outlier influence. Therefore, there exists a strong need for techniques that can reliably handle noise and outliers while still furnishing globally optimal solutions.
Proposed Methodologies
The authors propose a methodology grounded in semidefinite relaxations by introducing a truncated least squares cost function, aimed at enhancing robustness against outliers. Two principal formulations are discussed:
- Epipolar Constraint-Based Relaxation: This formulation capitalizes on lower dimensional constraints, ensuring efficacy under moderate noise levels. The method proves to be efficient in computation but begins to degrade with increased noise and outlier ratio.
- Fractional Reprojection Constraints-Based Relaxation: Alternatively, this formulation employs a higher dimensional space, thus offering better resilience against severe noise and prevalent outliers. Though computationally intensive, it manages to maintain tight relaxations even in adverse conditions.
Empirical validations through extensive simulation datasets reveal that these approaches deliver provably optimal reconstructions, demonstrating robustness even under significant noise and a considerable presence of outliers.
Results and Contributions
The paper provides strong numerical evidence of robustness with empirical validation using both synthetic and real-world datasets. The relaxations essentially guarantee optimal solutions in scenarios generally problematic for non-robust approaches. A notable achievement is maintaining tightness of the relaxation in the presence of both noise and outliers, thereby avoiding the degradation seen in prior methods.
Key contributions include:
- Extension of semidefinite relaxations for handling multiview triangulation with robust truncated least squares cost function.
- Design of two novel relaxation strategies accommodating a spectrum of noise and outlier levels.
- Empirical proofs confirming that the relaxations remain tight even under high disturbance conditions.
Implications and Future Directions
The implications of this research are broad, laying a foundation for applying semidefinite-relaxed formulations to other robust estimation problems in geometric vision. The proposed method can serve as a precursory tool for certifying local optimization methods by providing tight solutions, thus acting as a benchmark for further enhancement.
An avenue for future exploration involves the integration of these relaxation techniques with machine learning frameworks, potentially enhancing the robustness of deep learning-based vision systems. Another direction may explore scalability improvements, possibly through more efficient solver designs or distributed computing strategies, to accommodate more extensive datasets typical in real-time applications.
Overall, the paper enhances the toolbox available for robust computer vision systems, enabling more accurate triangulations even in less than optimal observational environments. This contribution is poised to impact fields relying heavily on precise triangulation, such as robotics, augmented and virtual reality, and autonomous navigation systems.
