Abstract
Visual tracking often fails when an object changes its proportions, such as a car turning or a person bending. This paper introduces IBCCF, a tracking framework that integrates a standard 2D center correlation filter with a family of four 1D boundary filters to independently track the target's edges. By enforcing a "near-orthogonality" constraint between these filters, the model achieves state-of-the-art robustness against aspect ratio variations.
Introduction
Most modern visual trackers rely on Correlation Filters (CF) because they are incredibly fast, leveraging the Fast Fourier Transform (FFT) to operate in the frequency domain. However, standard CF trackers have a major weakness: they typically assume the target's aspect ratio remains fixed.
While some "scale-adaptive" trackers can handle a target getting larger or smaller (isometric scaling), they struggle when the target deforms or rotates in a way that changes its height-to-width ratio. For example, a side-view of a car has a very different aspect ratio than a front-view. If a tracker cannot adjust its bounding box proportions, it eventually loses the target or includes too much background noise, leading to drift and failure.
Methodology: The IBCCF Framework
The authors propose the Integrating Boundary and Center Correlation Filters (IBCCF) model. The core innovation lies in how it rethinks the geometry of the target by decomposing the tracking task into center localization and boundary estimation.
1D Boundary Correlation Filters (BCFs)
Instead of relying solely on a 2D filter for the object's center, the authors introduce four 1D filters dedicated to localizing the left, right, top, and bottom boundaries. By treating boundaries as 1D signals, the model can flexibly "stretch" or "shrink" the bounding box in any direction independently. This allows the tracker to adapt to complex deformations that 2D filters often miss.
Near-Orthogonality Regularization
A key insight of this work is that the features representing the center of an object and its boundaries are naturally "nearly orthogonal"—meaning they provide non-overlapping, complementary information. The authors formalized this into a mathematical constraint (a regularization term) in their objective function. This forces the center and boundary filters to learn distinct, complementary features rather than redundant ones, ensuring that boundary filters focus on edges while the center filter focuses on the core appearance.
Optimization via ADMM
To solve the resulting complex optimization problem, the researchers employed the Alternating Direction Method of Multipliers (ADMM). This allows the model to iteratively update the center and boundary filters, typically converging to an optimal solution within five iterations.
Experimental Evaluation
The model was evaluated on several standard computer vision benchmarks, including OTB-2013, OTB-2015, Temple-Color, and the VOT series (2016/2017).
Key Results
- Superior Accuracy: On OTB-2015, IBCCF achieved a 78.4% mean Overlap Precision (OP), significantly outperforming its baseline counterpart, HCF (65.6%), and other scale-adaptive trackers like DSST and SAMF.
- Robustness to Rotation: In sequences involving in-plane and out-of-plane rotation—primary causes of aspect ratio changes—IBCCF showed gains of 2.5% and 6.4% respectively over standard deep-feature trackers.
- Efficiency: While more complex than a basic CF tracker, the ADMM approach remains relatively efficient, running at approximately 1.25 frames per second (fps) on a high-end PC (Xeon CPU + GTX 1080).
Discussion and Future Work
The most significant contribution of IBCCF is the "divide and conquer" strategy of moving from 2D to 1D for boundary detection. Enforcing near-orthogonality ensures the filters do not interfere with one another, leading to a 4.7% gain in precision on challenging sequences.
However, limitations remain. At 1.25 fps, the current implementation is not yet fast enough for real-time applications like drone navigation without further optimization. Future work may involve integrating advanced CF techniques, such as spatial regularization or continuous convolution (as seen in C-COT), to further enhance performance and speed.
Summary updated