Randomized Ensemble Tracking

Qinxun Bai, Zheng Wu, Stan Sclaroff, Margrit Betke, Camille MonnierICCV 2013 · DOI

Abstract

Visual tracking in non-stationary environments remains a significant challenge due to rapid changes in object appearance and frequent occlusions. This paper presents Randomized Ensemble Tracking (RET), a novel framework that treats classifier weights as random variables within a Bayesian estimation framework. By modeling weight distributions using a Dirichlet distribution and employing randomized voting, the proposed tracker dynamically adapts to changing conditions, effectively "forgetting" unreliable features while maintaining focus on visible target components.

Introduction: The Challenge of Non-Stationarity

Visual tracking-by-detection involves training an online classifier to distinguish a target from its background. However, traditional methods often struggle with non-stationarity—the phenomenon where an object's appearance evolves due to lighting shifts, rotations, or partial occlusions. Standard ensemble methods, such as online boosting, typically assume a fixed data distribution or rely on deterministic weight updates. These approaches can fail when a previously learned "positive" feature becomes misleading due to environmental changes. There is a critical need for tracking systems that can gracefully adapt their internal models by prioritizing currently reliable features over historical data.

Methodology: Randomized Ensemble Tracking (RET)

The core innovation of RET lies in its treatment of weak classifier reliability as a hidden state in a non-stationary system. The tracker decomposes the target into an 8x8 grid of local patches, each monitored by a weak classifier (a linear SVM).

Overview of the Randomized Ensemble Tracking (RET) framework showing the Bayesian update loop and randomized voting process.

Bayesian Weighting and Randomized Voting

Unlike deterministic models, RET treats the weight vector of these weak classifiers as a random variable following a Dirichlet distribution. At each frame, the system draws multiple samples (instantiations) from this distribution. Each sample generates a unique "strong classifier," and the final tracking decision is reached through a consensus vote among these randomized versions. This sampling process allows the tracker to explore model uncertainty, preventing it from prematurely converging on false positives or distractors.

Consistency-Based Updates

The model evolves through a recursive Bayesian update based on "consistency." If a weak classifier's prediction aligns with the final ensemble's consensus, its reliability score—and consequently its weight in the Dirichlet distribution—is increased. Conversely, if a classifier disagrees (e.g., when its corresponding patch is occluded), its influence is diminished.

Key Insights and Results

A primary advantage of RET is its inherent ability to handle partial occlusions without an explicit detection module. When a portion of the target is hidden, the classifiers for those specific patches begin to disagree with the visible majority, leading to a natural down-weighting of the occluded regions.

Demonstration of RET adapting to partial occlusion. The heatmap shows lower weights (blue) for the area occluded by the caliper.

Performance Evaluation

The researchers evaluated RET across 28 challenging benchmark sequences, including the Babenko, PROST, and VTD datasets. Key findings include:

  • State-of-the-Art Accuracy: RET achieved top-tier performance on 50% of the tested sequences, showing particular resilience in "distractor-heavy" environments where standard trackers often fail.
  • Robustness: The Bayesian weight evolution proved significantly more effective than simple error-rate updates used in standard online boosting.
  • Efficiency: The implementation maintains practical processing speeds, handling multiple scales in approximately 3 seconds per frame.

Limitations and Future Directions

While RET excels at feature adaptation, it currently lacks a formal motion model (e.g., a Kalman filter), meaning it does not utilize temporal dynamics like velocity or direction. Additionally, extremely similar distractors can still cause tracking drift due to a lack of global spatial context. Future work may involve integrating motion priors and exploring dynamic adaptation of the learning rate to further enhance robustness in highly complex scenes.

Summary updated

BibTeX
@inproceedings{bai2013randomized,
  author = {Qinxun Bai and Zheng Wu and Stan Sclaroff and Margrit Betke and Camille Monnier},
  title = {{Randomized Ensemble Tracking}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2013},
  doi = {10.1109/iccv.2013.255},
}
Figure