Source-linked AI summary
Large Margin Object Tracking with Circulant Feature Maps
Mengmeng Wang, Yong Liu, Zeyi Huang
TL;DR
Real-time visual tracking requires strong discrimination without the sampling and optimization costs of structured SVM methods. The paper proposes LMCF, which combines structured output SVM with correlation-filter acceleration, multimodal detection, and confidence-based model updates, and evaluates hand-crafted-feature and CNN versions. Across challenging benchmarks, LMCF and DeepLMCF outperform most compared state-of-the-art trackers while operating at real-time speeds.
Problem
Visual tracking must handle challenging target and background conditions in real time, while structured output SVM trackers are limited by time-consuming sampling and complex optimization.
Method
LMCF combines structured output SVM with dense circular samples and correlation-filter acceleration, adding multimodal detection, confidence-based updates, and hand-crafted-feature or CNN representations.
Results
LMCF performs best across OPE, TRE, and SRE metrics on OTB-13 and OTB-15, while DeepLMCF is second on OTB-13 precision and success and runs 40 times faster than C-COT.
Takeaways & Limitations
The evaluations support LMCF and DeepLMCF as fast tracking algorithms that perform well against most state-of-the-art methods using conventional and deep CNN features.
Abstract
from arXiv · showhide
Structured output support vector machine (SVM) based tracking algorithms have shown favorable performance recently. Nonetheless, the time-consuming candidate sampling and complex optimization limit their real-time applications. In this paper, we propose a novel large margin object tracking method which absorbs the strong discriminative ability from structured output SVM and speeds up by the correlation filter algorithm significantly. Secondly, a multimodal target detection technique is proposed to improve the target localization precision and prevent model drift introduced by similar objects or background noise. Thirdly, we exploit the feedback from high-confidence tracking results to avoid the model corruption problem. We implement two versions of the proposed tracker with the representations from both conventional hand-crafted and deep convolution neural networks (CNNs) based features to validate the strong compatibility of the algorithm. The experimental results demonstrate that the proposed tracker performs superiorly against several state-of-the-art algorithms on the challenging benchmark sequences while runs at speed in excess of 80 frames per second. The source code and experimental results will be made publicly available.
1. Introduction
Visual tracking must handle diverse target and background challenges in real time, but stronger discriminative models and richer features often increase computational cost. LMCF addresses this tension with dense circulant samples, correlation-filter acceleration, multimodal detection, and confidence-based updates.
- Real-time short-term tracking must handle occlusions, deformations, fast motion, illumination changes, and other target or background challenges.The paper frames tracking as single-object tracking with the target specified in the first frame.
- Existing trackers combine discriminative or generative models with hand-crafted or CNN features, but added algorithmic or representational complexity can reduce real-time performance.These approaches seek robust detection and model updates while trading computational efficiency against performance.
- Structured output SVM provides strong discrimination for tracking, yet Struck is limited by complex optimization and insufficiently dense training samples.The desired replacement should preserve structured-SVM discrimination while supporting higher-dimensional features and denser samples at sufficient speed.
- Correlation filters enable dense sampling and high-dimensional features in real time through FFT-based training and detection, while retaining computational efficiency.The paper also notes that conventional correlation filters can suffer boundary effects from their circular-shift assumption.
- LMCF is a structured-SVM tracker using dense circular samples in training and detection, linked to correlation filters to accelerate optimization.The method is presented as a large-margin object tracker with circulant feature maps.
- The proposed system adds multimodal target detection to reduce drift from similar objects or background noise and confidence-based updates to avoid model corruption.These mechanisms supplement the core tracker with localization and update safeguards.
2. Large Margin Object Tracking with Circulant Feature Maps
The paper develops a structured-output SVM tracker accelerated through circulant feature maps and Fourier-domain optimization. It adds multimodal detection and confidence-based updates to improve localization and prevent model corruption.
- Problem formulation: The tracker uses structured-output SVM to estimate relative target movement from cyclically shifted image patches rather than binary target labels.The formulation learns compatibility between input-output pairs and predicts by maximizing the response over candidate shifts.
- Fast online optimization: Dense circular samples are incorporated in training and detection, while Fourier transforms provide a faster optimization route related to correlation filtering.The optimization alternates closed-form updates for w and z, using DFT operations and element-wise multiplication.
- Multimodal target detection: Multimodal detection redetects regions around multiple response peaks when unimodal localization may select similar objects or background noise.The final target position is chosen from the maximum peak among the redetected response maps.
- High-confidence update: High-confidence feedback skips model updates when detections are unreliable, preventing uncertain samples from corrupting the tracking model.In severe occlusion, APCE falls to about 10 while Fmax remains strong, so the model is not updated and later tracking succeeds.
- High-confidence update: The APCE criterion measures response-map fluctuation and detection confidence, decreasing when the target is occluded or missing.APCE is larger for a smooth response map with one sharp peak and smaller for noisy or uncertain detections.
3. Experiments
Experiments evaluate LMCF and DeepLMCF on OTB benchmarks using standard robustness metrics, ablations, and comparisons with conventional- and CNN-feature trackers. LMCF provides the strongest conventional-feature results with high speed, while DeepLMCF improves accuracy and robustness over most CNN-based competitors.
- Evaluation setup: OTB-13 and OTB-15 evaluations use OPE, TRE, and SRE metrics across sequences annotated with 11 challenging attributes.The attributes include scale variation, occlusion, illumination variation, motion blur, deformation, fast motion, rotations, clutter, out-of-view, and low resolution.
- Ablation analysis: DeepLMCF achieves the best tracking accuracy and robustness among the LMCF variants, while LMCF ranks second with the fastest speed.The ablation compares multimodal detection, high-confidence updating, and hierarchical CNN features on OTB-13.
- Ablation analysis: Removing multimodal detection causes poor performance through false detections from similar objects or background noise.The variant is denoted LMCF-Uni.
- Ablation analysis: Removing high-confidence updating nearly halves tracking speed and lowers accuracy because LMCF-NU updates the model every frame.The comparison indicates that update frequency affects both operating efficiency and tracking accuracy.
- Evaluation on LMCF: LMCF performs best on OPE, TRE, and SRE across both OTB benchmarks, improving Struck by an average of 15% in average AUC scores.It also outperforms DSST and SAMF while running more than 3 times faster than DSST and more than 11 times faster than SAMF.
- Evaluation on LMCF: LMCF performs especially well under occlusion, scale variation, illumination variation, background clutter, and out-of-plane rotation.These are among the challenging attributes analyzed on OTB-13.
- Evaluation on DeepLMCF: DeepLMCF ranks second behind C-COT on precision and success scores, exceeds 10 FPS, and outperforms the other compared CNN trackers except C-COT.Its reported speed is 40 times faster than C-COT, whose reported speed is about 0.25 FPS.
4. Conclusion
The paper combines large-margin discrimination with correlation-filter efficiency, multimodal detection, and high-confidence updates. LMCF and DeepLMCF show strong benchmark performance, with DeepLMCF demonstrating CNN compatibility.
- LMCF links structured output SVM discrimination with correlation-filter efficiency for fast object tracking.The method uses a bridge between the large-margin formulation and correlation filters.
- Multimodal target detection addresses localization errors and model drift from similar objects or background noise.
- High-confidence model updates use tracking-result feedback to avoid model corruption.
- LMCF performs best in almost all eight challenging attributes evaluated in Figure 4.The attributes include clutter, illumination variation, occlusion, deformation, rotations, out-of-view targets, and scale variation.
- Figure 5 reports OTB-13 OPE precision and success plots, with legend values representing average precision and average AUC scores.
- DeepLMCF extends the tracker with deep CNN features, demonstrating compatibility across feature representations.