DeepTrack: Learning Discriminative Feature Representations Online for Robust Visual Tracking (1503.00072v1)

Abstract: Deep neural networks, albeit their great success on feature learning in various computer vision tasks, are usually considered as impractical for online visual tracking because they require very long training time and a large number of training samples. In this work, we present an efficient and very robust tracking algorithm using a single Convolutional Neural Network (CNN) for learning effective feature representations of the target object, in a purely online manner. Our contributions are multifold: First, we introduce a novel truncated structural loss function that maintains as many training samples as possible and reduces the risk of tracking error accumulation. Second, we enhance the ordinary Stochastic Gradient Descent approach in CNN training with a robust sample selection mechanism. The sampling mechanism randomly generates positive and negative samples from different temporal distributions, which are generated by taking the temporal relations and label noise into account. Finally, a lazy yet effective updating scheme is designed for CNN training. Equipped with this novel updating algorithm, the CNN model is robust to some long-existing difficulties in visual tracking such as occlusion or incorrect detections, without loss of the effective adaption for significant appearance changes. In the experiment, our CNN tracker outperforms all compared state-of-the-art methods on two recently proposed benchmarks which in total involve over 60 video sequences. The remarkable performance improvement over the existing trackers illustrates the superiority of the feature representations which are learned

• The paper introduces a truncated structural loss function to prioritize key training samples and mitigate error accumulation.
• The paper advances online CNN training with enhanced SGD and temporal sampling, achieving superior tracking precision and robustness on benchmarks.
• The paper employs a lazy updating scheme that adapts to object appearance changes, ensuring real-time performance and reliability.

Overview of DeepTrack: A CNN-Based Approach for Online Visual Tracking

The paper introduces "DeepTrack," a visual tracking algorithm leveraging convolutional neural networks (CNNs) to learn discriminative feature representations in a purely online setting. This approach addresses common challenges in visual tracking by employing CNNs, traditionally seen as impractical due to significant training data and computational demands.

Key Contributions

DeepTrack proposes several novel mechanisms to improve the robustness and efficiency of online visual tracking using CNNs:

1. Truncated Structural Loss Function: The authors introduce a truncated structural loss function to manage more training samples while mitigating error accumulation. This function strategically prioritizes samples with meaningful errors, reducing computational overhead and promoting efficient learning.
2. Enhanced Stochastic Gradient Descent (SGD): An advanced SGD approach is employed, integrating a temporal sampling mechanism that diversifies training data by considering temporal relationships and label noise. This method helps in regularizing CNN training and mitigating overfitting.
3. Lazy Updating Scheme: The CNN model is updated in a lazy yet effective manner. This scheme primarily triggers updates in response to significant changes in object appearance, tackling challenges like occlusion or incorrect detections without compromising adaptation capabilities.

Strong Numerical Results and Benchmark Comparisons

Evaluations on two recently proposed benchmarks with over 60 video sequences highlight DeepTrack's superior performance. The tracker outperformed state-of-the-art methods significantly:

• CVPR2013 Benchmark: DeepTrack achieved an 83% tracking precision and a 63% success rate, outperforming the closest competitors (TGPR, KCF) by substantial margins.
• VOT2013 Benchmark: DeepTrack ranked first across accuracy and robustness metrics compared to 27 other trackers, demonstrating superior balance in tracking accuracy and fail-safe robustness.

Implications and Future Research Directions

The DeepTrack algorithm illustrates the efficacy of using CNNs for online visual tracking, providing insights into the real-time application of deep learning models:

• Practical Implications: The robust design of DeepTrack, with efficiency measures like lazy updating and temporal sampling, make it viable for practical applications requiring real-time processing.
• Theoretical Implications: The paper's contributions, particularly the truncated structural loss and enhanced training dynamics, offer a template for reducing the computational burden in deep learning while maintaining robustness.
• Future Directions: The exploration of additional cues and the extension of this method to handle more complex or dynamic environments could lead to further advancements in visual tracking. Additionally, research can explore integrating this approach with fully unsupervised learning paradigms to enhance adaptability.

Overall, the paper provides a comprehensive framework for employing CNNs in online visual tracking, addressing existing constraints and setting a foundation for future developments in this area.