Learning Support Correlation Filters for Visual Tracking

I. INTRODUCTION

• SCF-based methods exploit circulant dense sampling to address SVM trackers’ sampling and budgeting trade-off, with O(n^2 log n) complexity versus at least O(n^4) for standard SVM approaches.The paper also extends SCF to multi-channel, kernelized, and scale-adaptive variants.Source: §I · p. 1
• KSCF and SKSCF perform favorably against state-of-the-art trackers in accuracy and speed, while KSCF also compares favorably with regression-based correlation filters and SVM-based trackers.The reported comparisons use a large benchmark dataset.Source: §I · p. 1

A. Appearance models for visual tracking

• Discriminative tracking uses classifiers including boosting, random forests, and SVMs, while this work combines SVM discriminative strength with correlation-filter efficiency.The paper also addresses label ambiguity using an assignment scheme similar to those used in object detection and tracking.Source: A. Appearance models for visual tracking · p. 2

B. Correlation filters for tracking

• Earlier correlation-filter trackers generally use ridge-regression predictors, while combining SVMs with the max-margin correlation-filter approach is computationally prohibitive for real-time tracking.The paper positions its approach as an efficient SVM-and-correlation-filter alternative.Source: B. Correlation filters for tracking · p. 2

A. Problem formulation

• Translated image patches form a circulant matrix; confidence-map thresholds assign positive and negative labels while leaving uncertain samples unlabeled for training.This assignment is used to alleviate label ambiguity.Source: A. Problem formulation · pp. 2–3
• The SCF formulation learns a max-margin classifier and bias to distinguish target from background, rather than fitting a confidence map as regression-based correlation filters do.The comparison is illustrated in Fig. 2.Source: A. Problem formulation · p. 3, Figure 2 · p. 4

B. Alternating optimization

• The algorithm alternates between updating e and updating {w, b}, with closed-form subproblems enabling efficient optimization.Given {w, b}, e is updated element-wise; given e, the classifier and bias are updated through a quadratic-program subproblem.Source: B. Alternating optimization · p. 4
• Each iteration has O(n^2 log n) complexity, with O(n^2) updates for e and b and an O(n^2 log n) update for w.The algorithm iterates until convergence.Source: B. Alternating optimization · p. 4
• The proposed algorithm converges to the global optimum with a q-linear convergence rate.The paper says its detailed optimality and convergence analysis is provided in Appendix A.Source: B. Alternating optimization · p. 4
• For tracking with N=n^2 samples and dimension d=n^2, SCF has O(n^2 log n) complexity, compared with MMCF’s O(n^6) SMO complexity.SCF uses squared hinge loss and regularizer terms; MMCF uses hinge loss and an extra average correlation energy term.Source: B. Alternating optimization · p. 5

C. Multi-channel SCF

• The multi-channel filter uses L image channels and corresponding per-channel image and filter components, x_l and w_l.The approach treats local descriptors as multi-channel images, with multiple measurements associated with each pixel.Source: C. Multi-channel SCF · p. 5
• The channel-wise structure decomposes the filter update into n^2 systems, each depending only on the corresponding frequency components across channels.The passage contrasts this decomposition with directly computing the inverse of the full matrix.Source: C. Multi-channel SCF · p. 5
• The proposed update uses a rank-one matrix plus a scaled identity and the Sherman-Morrison formula to obtain a closed-form channel solution.This is presented after comparison with an existing solver for the n^2 systems.Source: C. Multi-channel SCF · p. 5
• Each iteration requires one DFT, one IDFT, and element-wise operations, giving complexity O(n^2 log n).The passage notes that the Fourier-domain quantities can be precomputed at O(n^2 log n) complexity.Source: C. Multi-channel SCF · p. 5

D. Kernelized SCF

• The kernelized SCF model extends the decision function to nonlinear feature mappings defined implicitly by a kernel K(x, x′).The dual representation uses a learned coefficient vector α over the samples.Source: D. Kernelized SCF · p. 5
• For permutation-invariant kernels such as Gaussian RBF and polynomial kernels, the circulant kernel matrix permits matrix-vector multiplication via DFT.The passage identifies the kernel matrix's first row as the kernel correlation used in this computation.Source: D. Kernelized SCF · p. 5
• KSCF learns by alternating updates of e and {w, b}, with the latter subproblem solved using a closed-form kernel ridge regression solution.The kernel-space solution is represented as a linear combination of sample feature maps.Source: D. Kernelized SCF · p. 5
• For L-channel image features, kernel computation costs O(Ln^2 log n), and each learning iteration costs O(n^2 log n).The passage says this allows nonlinear filters to use rich features without significantly increasing computational load.Source: D. Kernelized SCF · p. 6
• SKSCF handles large scale changes with a scaling pool and uses bilinear interpolation to resize samples across scales.Source: D. Kernelized SCF · p. 6

IV. PERFORMANCE EVALUATION

• The evaluation compares KSCF and SKSCF with existing SVM-based and state-of-the-art methods.Source: §IV · p. 6
• The experiments assess SCF, MSCF, KSCF, and SKSCF variants to analyze feature representations and kernel functions.Source: §IV · p. 6
• Comprehensive experiments compare the proposed methods with other correlation-filter-based trackers.Source: §IV · p. 6

A. Experimental setup

• The benchmark comprises 50 challenging sequences annotated with 11 attributes, and evaluation uses OPE precision and success plots plus FPS.Precision reports distance precision at 20 pixels; success reports AUC, with ground-truth locations used for position error and bounding-box overlap.Source: A. Experimental setup · p. 6

B. Evaluation on SCF-based trackers

• HOG and color-name features improve mean DP over raw pixels by 13.5% and 1.4% for MSCF, and by 14.9% and 3.7% for KSCF, respectively.With HOG and color names combined, MSCF reaches 80.6% DP and KSCF reaches 85.0% DP; both remain real-time with HOG and CN features.Source: B. Evaluation on SCF-based trackers · p. 7, Figure 3 · p. 7, Figure 4 · p. 7
• The Gaussian RBF kernel gives KSCF its best performance among the tested kernels, outperforming the linear kernel in mean DP and mean AUC.The Gaussian RBF parameter σ is set to 0.2.Source: B. Evaluation on SCF-based trackers · p. 7, Table III · p. 7
• Among the four SCF variants, SKSCF performs best, followed by KSCF, while KSCF is more efficient than SKSCF.SKSCF extends KSCF with the Gaussian RBF kernel.Source: B. Evaluation on SCF-based trackers · p. 7

C. Comparisons with CF-based trackers

• KSCF using HOG and color names performs favorably against KCF and CSK, while MSCF performs well among the compared multi-channel correlation filters.MSCF combines HOG and color representations; DCF uses HOG and CN uses color attributes.Source: C. Comparisons with CF-based trackers · pp. 8–9
• KSCF and SKSCF significantly outperform STC, DSST, and SAMF, with KSCF favorable on scale-variation sequences and in accuracy and speed overall.DSST and SAMF had been reported among the strongest trackers in a recent benchmark evaluation.Source: C. Comparisons with CF-based trackers · p. 9

D. Comparisons with SVM-based trackers

• KSCF tracks objects more precisely than the compared SVM-based methods in challenging sequences, where other trackers tend to drift.The reported examples include Singer2, Coke, Suv, and Tiger2.Source: D. Comparisons with SVM-based trackers · p. 9

E. Comparisons with state-of-the-art trackers

• KSCF and SKSCF perform favorably against state-of-the-art trackers, including TLD, SCM, TGPR, and MEEM.The comparison reports OPE plots and mean DP, AUC, and FPS.Source: E. Comparisons with state-of-the-art… · p. 9, §V · p. 10
• SKSCF performs best in both metrics for all but one attribute, while KSCF ranks among the top three for every attribute.MEEM performs slightly better than KSCF on FM, LR, OV, and SV.Source: E. Comparisons with state-of-the-art… · p. 10, Table VII · p. 10, Table VIII · p. 10

V. CONCLUSIONS

• The paper concludes that KSCF and SKSCF perform favorably against state-of-the-art trackers in accuracy and speed.The proposed variants include multi-channel, nonlinear, and scale-adaptive tracking methods.Source: §V · p. 10

• The appendix analyzes the SCF algorithm’s optimality condition, global convergence, and convergence rate.Source: Appendix A · p. 10

A. Optimality conditions

• For convex loss (22), (˜w∗, e∗) is a solution if and only if the loss subdifferential at that point contains zero.Source: A. Optimality conditions · p. 11

B. Global convergence

• The analysis characterizes fixed points of g ◦h and uses them to establish convergence of the error sequence.The stated lemma says that distance to a fixed point decreases unless the iterate is itself a fixed point.Source: B. Global convergence · p. 11
• From any starting point (˜w0, e0), the algorithm’s sequence converges to a solution of the optimization problem.The proof identifies the limit as a fixed point satisfying the optimization conditions.Source: B. Global convergence · p. 12
• The sequence ˜w_k converges to ˜w∗ q-linearly.Source: C. q-linear convergence rate · p. 12