Abstract:While automatically generated polynomial elimination templates have sparked great progress in the field of 3D computer vision, there remain many problems for which the degree of the constraints or the number of unknowns leads to intractability. In recent years, homotopy continuation has been introduced as a plausible alternative. However, the method currently depends on expensive parallel tracking of all possible solutions in the complex domain, or a classification network for starting problem-solution pairs trained over a limited set of real-world examples. Our innovation consists of employing a regression network trained in simulation to directly predict a solution from input correspondences, followed by an online simulator that invents a consistent problem-solution pair. Subsequently, homotopy continuation is applied to track that single solution back to the original problem. We apply this elegant combination to generalized camera resectioning, and also introduce a new solution to the challenging generalized relative pose and scale problem. As demonstrated, the proposed method successfully compensates the raw error committed by the regressor alone, and leads to state-of-the-art efficiency and success rates while running on CPU resources, only.

What problem does this paper attempt to address?

### Problems the paper attempts to solve This paper aims to solve the problem of solving complex polynomial equations in geometric vision problems, especially in camera resectioning and generalized relative pose and scale estimation. Specifically: 1. **Limitations of existing methods**: - Traditional polynomial elimination - based methods (such as Gröbner basis method and polynomial resultant method) face challenges in computational complexity and numerical stability when dealing with high - dimensional or complex - constraint problems. - Although the Homotopy Continuation (HC) method is effective, it needs to track all possible solutions in parallel, resulting in high computational cost. 2. **Proposed new method**: - The author introduces a new framework that combines a regression network and an online simulator to generate consistent initial problem - solution pairs and tracks a single solution through the Homotopy Continuation method. - This method can significantly improve the solution efficiency and can run efficiently on the CPU, being suitable for more complex geometric problems. 3. **Specific applications**: - **Generalized camera resectioning**: Solves the problem of estimating the camera pose from 2D - 3D correspondences. - **Generalized relative pose and scale estimation**: Solves the problem of estimating the relative pose and scale between two cameras from 2D - 2D correspondences, which is a challenging minimization problem. ### Core innovation points of the method - **Regression network**: Used to predict an approximate solution from the input correspondences instead of relying on a limited real - world data set for classification. - **Online simulator**: Generates a new problem - solution pair consistent with the predicted solution according to the prediction of the regression network. - **Single - root tracking**: Improves computational efficiency by tracking only one solution through the Homotopy Continuation method. ### Results and advantages - **High efficiency**: Compared with the traditional HC method, the new method is more efficient when running on the CPU, with a success rate of over 95%. - **Generality**: This method is not only applicable to simple geometric problems, but also can handle more complex generalized relative pose and scale estimation problems. - **Accuracy**: Although the solution predicted by the regression network may not be accurate enough, it provides a good initial solution for Homotopy Continuation and finally obtains accurate results. In conclusion, by combining deep learning and the Homotopy Continuation method, this paper provides a novel and efficient solution that can find accurate solutions in complex geometric problems while maintaining high computational efficiency.

Homotopy Continuation Made Easy: Regression-based Online Simulation of Starting Problem-Solution Pairs

MCNet: Rethinking the Core Ingredients for Accurate and Efficient Homography Estimation

A Compensatory Algorithm for High-Speed Visual Object Tracking Based on Markov Chain

Continuation Path Learning for Homotopy Optimization

Homography Decomposition Networks for Planar Object Tracking

Backpropagation through Combinatorial Algorithms: Identity with Projection Works

Online Stability Improvement of Groebner Basis Solvers using Deep Learning

On the convergence of graph matching: graduated assignment revisited

Every Pixel Counts ++: Joint Learning of Geometry and Motion with 3D Holistic Understanding

Generating 3D-Consistent Videos from Unposed Internet Photos

RecON: Online learning for sensorless freehand 3D ultrasound reconstruction

Memory-efficient model-based deep learning with convergence and robustness guarantees

View Extrapolation of Human Body from a Single Image

Two Sides of The Same Coin: Bridging Deep Equilibrium Models and Neural ODEs via Homotopy Continuation

Optimization problem in multi-homogeneous homotopy method

HomPINNs: Homotopy physics-informed neural networks for solving the inverse problems of nonlinear differential equations with multiple solutions

SOLVING NONLINEAR COMPLEMENTARITY PROBLEM BY A SMOOTHING HOMOTOPY METHOD

Monodromy Solver: Sequential and Parallel

Every Pixel Counts: Unsupervised Geometry Learning with Holistic 3D Motion Understanding

Content-Aware Unsupervised Deep Homography Estimation and its Extensions