Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

OCP-LS: An Efficient Algorithm for Visual Localization

Jindi Zhong, Hongxia Wang, Huanshui Zhang

TL;DR

This paper addresses efficient and robust training for visual localization by modeling CNN optimization as a discrete-time optimal-control problem. The proposed OCP-LS algorithm uses a Gauss-Newton diagonal approximation via a GNB-inspired estimator and a Newton-like update to guide the parameter trajectory. Theoretical guarantees of linear convergence under standard assumptions are provided, and extensive experiments on Cambridge landmark datasets demonstrate faster convergence and robust performance with competitive localization accuracy. The work offers a system-level perspective on training dynamics for visual localization, bridging optimal control and deep learning methodologies.

Abstract

This paper proposes a novel second-order optimization algorithm. It aims to address large-scale optimization problems in deep learning because it incorporates the OCP method and appropriately approximating the diagonal elements of the Hessian matrix. Extensive experiments on multiple standard visual localization benchmarks demonstrate the significant superiority of the proposed method. Compared with conventional optimiza tion algorithms, our framework achieves competitive localization accuracy while exhibiting faster convergence, enhanced training stability, and improved robustness to noise interference.

OCP-LS: An Efficient Algorithm for Visual Localization

TL;DR

This paper addresses efficient and robust training for visual localization by modeling CNN optimization as a discrete-time optimal-control problem. The proposed OCP-LS algorithm uses a Gauss-Newton diagonal approximation via a GNB-inspired estimator and a Newton-like update to guide the parameter trajectory. Theoretical guarantees of linear convergence under standard assumptions are provided, and extensive experiments on Cambridge landmark datasets demonstrate faster convergence and robust performance with competitive localization accuracy. The work offers a system-level perspective on training dynamics for visual localization, bridging optimal control and deep learning methodologies.

Abstract

This paper proposes a novel second-order optimization algorithm. It aims to address large-scale optimization problems in deep learning because it incorporates the OCP method and appropriately approximating the diagonal elements of the Hessian matrix. Extensive experiments on multiple standard visual localization benchmarks demonstrate the significant superiority of the proposed method. Compared with conventional optimiza tion algorithms, our framework achieves competitive localization accuracy while exhibiting faster convergence, enhanced training stability, and improved robustness to noise interference.
Paper Structure (5 sections, 1 theorem, 19 equations, 1 figure, 1 table, 1 algorithm)

This paper contains 5 sections, 1 theorem, 19 equations, 1 figure, 1 table, 1 algorithm.

Table of Contents

  1. Problem Formulation
  2. The Optimization Algorithm
  3. GNB estimator
  4. Update Rules
  5. Experimetal Results

Key Result

Theorem 1

Suppose that Assumption ass:main holds. The proposed algorithm ocp-ls generates a sequence $\{x_k\}$ that converges to the optimum $x^{*}$ asymptotically at a linear rate. Specifically, there exists a constant $\rho_{\infty} \in [0,1)$ such that where $\rho_{\infty}$ represents the asymptotic linear convergence rate of the algorithm.

Figures (1)

  • Figure 1: Training and Validation Loss curves of the Cambridge Landmarks dataset

Theorems & Definitions (1)

  • Theorem 1