Object Tracking Paper(8): IBCCF-- Integrating Boundary and Center Correlation Filters for VOT

The seventh paper: IBCCF--Integrating Boundary and Center Correlation Filters for Visual Tracking with Aspect Ratio Variation/ Author: Feng Li, Yingjie Yao, Peihua Li, David / Publication information: CVPR 2017

Outline: This paper presents a framework which combine 2D CF and 1D boundary CF to solve the problem of aspect ratio variation. They discover the near-orthogonality between the center and the four boundaries. And they add the near-orthogonality regularization into the loss function. Obviously, it is not a convex problem, so they deploy the ADMM(Alternating Direction Method of Multipliers) to solve this long regression problem. There are six variables to be updated to get the final solution. They update and find the close-form solution of one variable and fix other variables. During finding the close-form solution, they use Singular Value Decomposition and Sherman-Morrison formula to solve the inverse matrix problem. They conduct plenty experiments to prove that this work can truly solve the problem of aspect ratio variation, though the performance is not good as C-COT.

Methodology: 1. Features: CNN features from VGG Net. And the features are very similar to HCF. They compare the HCF with scale variation and this work.

2. The near-orthogonality, ADMM, The augmented lagrangian form and Sherman-Morrison formula

Advantages: They prove the near-orthogonality property of center and the boundary. This regularization guarantees the favorable performance.
Disadvantages: 1.25 fps on CPU, the CF of Center and four boundaries are redundant, it aggregates the computational burden.