Zero-Aliasing Correlation Filters for Object Recognition

Joseph A. Fernandez, Vishnu Naresh Boddeti, Andres Rodriguez, B. V. K. Vijaya Kumar

Abstract

Correlation filters (CFs) are a class of classifiers that are attractive for object localization and tracking applications. Traditionally, CFs have been designed in the frequency domain using the discrete Fourier transform (DFT), where correlation is efficiently implemented. However, existing CF designs do not account for the fact that the multiplication of two DFTs in the frequency domain corresponds to a circular correlation in the time/spatial domain. Because this was previously unaccounted for, prior CF designs are not truly optimal, as their optimization criteria do not accurately quantify their optimization intention. In this paper, we introduce new zero-aliasing constraints that completely eliminate this aliasing problem by ensuring that the optimization criterion for a given CF corresponds to a linear correlation rather than a circular correlation. This means that previous CF designs can be significantly improved by this reformulation. We demonstrate the benefits of this new CF design approach with several important CFs. We present experimental results on diverse data sets and present solutions to the computational challenges associated with computing these CFs.

Introduction: The Aliasing Problem in Correlation Filters

Correlation filters are widely used for object localization due to their computational efficiency. By leveraging the Discrete Fourier Transform (DFT), spatial correlation—a computationally intensive operation—is transformed into simple element-wise multiplication in the frequency domain.

However, this efficiency comes with a significant mathematical caveat: the DFT assumes signals are periodic. Consequently, multiplying two DFTs in the frequency domain results in a circular correlation in the spatial domain, rather than the linear correlation typically desired for real-world testing. This discrepancy leads to "aliasing," where the edges of the correlation result wrap around and interfere with the signal.

Comparison of (a) Input Signal, (b) Linear Correlation, and (c) Circular Correlation showing the effects of aliasing.

While practitioners often use zero-padding during the testing phase to mitigate these effects, the aliasing has historically been ignored during the filter design (training) phase. This means traditional CFs are optimized for a circular environment that does not match actual deployment conditions, resulting in noisy outputs and less distinct detection peaks.

Methodology: Zero-Aliasing Correlation Filters (ZACFs)

To address this fundamental flaw, the authors propose Zero-Aliasing Correlation Filters (ZACFs). The core innovation is the introduction of spatial constraints that force the "tail" of the filter template to zero.

The ZACF pipeline compared to conventional CFs, highlighting the tail constraint that eliminates aliasing.

The ZACF framework involves:

  1. Zero-Aliasing (ZA) Constraints: A constraint matrix is mathematically defined to force filter coefficients to zero outside the intended template size. This ensures the optimization criterion (such as Mean Squared Error) accurately reflects linear correlation.
  2. Reformulating Classic Filters: This logic is applied to well-known filters, resulting in "Zero-Aliasing" versions of MACE, MOSSE, OTSDF, and MMCF.
  3. Efficient Optimization via Proximal Gradient Descent: Since solving these new equations exactly is computationally expensive, the authors utilize an Accelerated Proximal Gradient Descent method. This iterative approach allows the "zero-tail" constraint to be imposed by simply cropping the filter in the spatial domain at each step, making the training of large filters feasible.

Experimental Results

The ZACF approach was validated across diverse tasks, including 1-D signal analysis (ECG), face recognition (AT&T and FRGC datasets), target recognition (infrared vehicle images), eye localization (FERET), and general object detection (INRIA Pedestrians and ETHZ Shapes).

Key findings include:

  • Improved Accuracy: In face recognition on the FRGC dataset, the ZAMACE filter improved the Rank-1 Identification rate from 52.73% to 78.98% compared to the standard MACE filter.
  • Precision in Localization: For eye localization, ZACFs (ZAMOSSE and ZAMMCF) achieved higher accuracy at strict thresholds (e.g., 97% vs 94%), demonstrating better resistance to distractions from nearby facial features.
  • Pedestrian Detection: On the INRIA dataset, the ZAMMVCF achieved an Average Precision of 84.07%, outperforming standard vector correlation filters.

Key Insights and Conclusion

A critical takeaway is that simply zero-padding input images during training is insufficient; the filter itself must be constrained to truly eliminate aliasing. By bridging the spatial and frequency domains—calculating gradients in the frequency domain for speed while projecting constraints in the spatial domain for accuracy—ZACFs produce sharper detection peaks and higher Peak-to-Correlation Energy (PCE).

While ZACFs require more training time than traditional CFs, their testing speed remains identical, making them a powerful drop-in replacement for localization tasks. Future work may extend these zero-aliasing constraints to more complex architectures like Fourier-domain Convolutional Neural Networks or non-linear kernels.

Code for the CFs described is available at http://vishnu.boddeti.net/projects/correlation-filters.html.

Summary updated

Abstract

Correlation filters (CFs) are a class of classifiers that are attractive for object localization and tracking applications. Traditionally, CFs have been designed in the frequency domain using the discrete Fourier transform (DFT), where correlation is efficiently implemented. However, existing CF designs do not account for the fact that the multiplication of two DFTs in the frequency domain corresponds to a circular correlation in the time/spatial domain. Because this was previously unaccounted for, prior CF designs are not truly optimal, as their optimization criteria do not accurately quantify their optimization intention. In this paper, we introduce new zero-aliasing constraints that completely eliminate this aliasing problem by ensuring that the optimization criterion for a given CF corresponds to a linear correlation rather than a circular correlation. This means that previous CF designs can be significantly improved by this reformulation. We demonstrate the benefits of this new CF design approach with several important CFs. We present experimental results on diverse data sets and present solutions to the computational challenges associated with computing these CFs. Code for the CFs described in this paper and their respective zero-aliasing versions is available at http://vishnu.boddeti.net/projects/correlation-filters.html
BibTeX
@article{fernandez2014zero,
  author = {Joseph A. Fernandez and Vishnu Naresh Boddeti and Andres Rodriguez and B. V. K. Vijaya Kumar},
  title = {{Zero-Aliasing Correlation Filters for Object Recognition}},
  journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year = {2014},
  doi = {10.1109/tpami.2014.2375215},
}
Figure