Learning Spatially Regularized Correlation Filters for Visual Tracking (1608.05571v1)

Abstract: Robust and accurate visual tracking is one of the most challenging computer vision problems. Due to the inherent lack of training data, a robust approach for constructing a target appearance model is crucial. Recently, discriminatively learned correlation filters (DCF) have been successfully applied to address this problem for tracking. These methods utilize a periodic assumption of the training samples to efficiently learn a classifier on all patches in the target neighborhood. However, the periodic assumption also introduces unwanted boundary effects, which severely degrade the quality of the tracking model. We propose Spatially Regularized Discriminative Correlation Filters (SRDCF) for tracking. A spatial regularization component is introduced in the learning to penalize correlation filter coefficients depending on their spatial location. Our SRDCF formulation allows the correlation filters to be learned on a significantly larger set of negative training samples, without corrupting the positive samples. We further propose an optimization strategy, based on the iterative Gauss-Seidel method, for efficient online learning of our SRDCF. Experiments are performed on four benchmark datasets: OTB-2013, ALOV++, OTB-2015, and VOT2014. Our approach achieves state-of-the-art results on all four datasets. On OTB-2013 and OTB-2015, we obtain an absolute gain of 8.0% and 8.2% respectively, in mean overlap precision, compared to the best existing trackers.

• The paper introduces spatial regularization in DCFs to overcome boundary effects and improve the discriminative power of trackers.
• It utilizes an efficient Fourier domain optimization with the iterative Gauss-Seidel method to handle the complex optimization problem.
• Experiments on multiple benchmarks reveal significant accuracy gains, establishing state-of-the-art performance in visual tracking.

Learning Spatially Regularized Correlation Filters for Visual Tracking

The paper "Learning Spatially Regularized Correlation Filters for Visual Tracking" by Danelljan et al. introduces a novel approach to improve the robustness and accuracy of visual trackers by enhancing the discrimination capability of correlation filters. The primary innovation lies in the introduction of spatial regularization within the discriminative correlation filter (DCF) framework.

Key Contributions

The DCF methodology has been effective in visual tracking due to its efficiency and effectiveness in training models from limited training samples. However, the periodic assumption inherent in DCFs introduces boundary effects which degrade the quality and robustness of the learned appearance model. To address these limitations, the authors propose Spatially Regularized Discriminative Correlation Filters (SRDCF).

Spatial Regularization in DCF

The SRDCF incorporates a spatial regularization component that penalizes the filter coefficients based on their spatial location, thus allowing the learning of correlation filters over larger image regions. This mitigates the boundary effects seen in traditional DCFs by enabling the expansion of the training image region without incorporating background noise into the positive samples. Consequently, a more discriminative target appearance model is constructed.

Theoretical Foundations and Optimization

The incorporation of spatial regularization complicates the optimization problem. The authors derive an efficient optimization strategy in the Fourier domain, utilizing the iterative Gauss-Seidel method to solve the resulting normal equations. This approach leverages the sparsity of the proposed regularizer, ensuring computational efficiency suitable for online learning environments typical in tracking scenarios.

Experimental Evaluation

Extensive experimentation was conducted on four benchmark datasets: OTB-2013, OTB-2015, ALOV++, and VOT2014. The SRDCF achieved state-of-the-art performance across all datasets. Specifically, the proposed method obtained absolute gains of 8.0% on OTB-2013 and 8.2% on OTB-2015 in mean overlap precision over the best existing trackers. Moreover, SRDCF outperformed the top-ranked trackers in terms of both accuracy and robustness in VOT2014.

Implications and Future Directions

The introduction of spatial regularization advances the state of DCF-based trackers by addressing one of their main limitations. This improvement enables DCFs to handle complex scenarios such as target deformation, occlusions, fast motion, and scale variations more effectively. The enhanced discriminative power of the SRDCF model suggests potential in other areas where correlation filters are applied, including object detection and recognition.

Future research may explore adaptive forms of spatial regularization that dynamically adjust to changing target characteristics or leverage more sophisticated regularizers that incorporate spatial and temporal information. Additionally, integrating deep learning features into the SRDCF framework could further enhance its robustness and accuracy.

Conclusion

The paper by Danelljan et al. makes significant strides in improving the efficacy of visual trackers by addressing the boundary effects in DCFs through spatial regularization. The SRDCF framework not only achieves superior performance across multiple benchmarks but also sets a solid foundation for future research in visual tracking and beyond.