Online Visual Tracking using Multiple Instance Learning with Instance Significance Estimation

Tianfei Zhou, Yao Lu

Abstract

Visual tracking-by-detection often encounters the "drifting" problem caused by label ambiguity during online updates. This paper proposes an Instance Significance Guided Multiple Instance Learning (MIL) framework to address this. By estimating the "significance" of individual instances within a positive bag using a Bayesian ensemble approach, the tracker becomes more robust to noise and motion blur. Experimental results on nine challenging sequences demonstrate that the proposed method outperforms several state-of-the-art trackers in both accuracy and overlap rate.

Introduction

In the "tracking-by-detection" paradigm, a classifier is updated online to distinguish the target from the background. A major challenge is label ambiguity: if only the current tracking result is used as a positive sample, small localization errors accumulate, causing the tracker to "drift" onto the background. While Multiple Instance Learning (MIL) addresses this by using "bags" of samples, standard MIL assumes all instances in a positive bag are equally likely to be the target. This paper argues that treating near-misses and true targets identically introduces noise that degrades performance, especially during rapid motion or occlusion.

Instance Significance Guided MIL

The authors introduce Instance Significance Guided MIL Boosting, which improves upon traditional MIL through two key innovations:

1. Significance-Coefficient Estimation

Rather than relying on a single classifier, the system trains an ensemble of $N$ randomized MILBoost learners. A Bayesian formula is used to calculate a "significance-coefficient" ($r_{ij}$) for each instance. This coefficient reflects the likelihood that a specific patch represents the true target based on the collective agreement of the ensemble. Unlike distance-based weighting, this discriminative approach identifies high-quality samples even when the tracker has begun to drift.

2. Refined MILBoost Model

The traditional "Noisy-OR" model, which determines the probability of a bag being positive, is modified to include these significance coefficients as exponents in the likelihood function. This ensures that instances with high significance have a dominant impact on the model update, while noisy or ambiguous patches are effectively downweighted.

Experimental Evaluation

The proposed tracker was evaluated on nine public video sequences (over 9,000 frames) featuring challenges such as motion blur (Boy, Doll), low resolution (Panda), and illumination changes (Dollar).

Quantitative Results

The method was compared against six state-of-the-art trackers, including ASLA, PCOM, and the original MIL. As shown in the performance table below, the proposed method achieved the best average Center Location Error (CLE) of 7.8 pixels and the highest average VOC Overlap Rate (VOR) of 0.75.

Quantitative comparison of tracking performance across nine datasets using CLE and VOR metrics.

Qualitative Analysis

Qualitative results demonstrate the tracker's robustness. While competing methods often failed during long-term appearance changes or significant motion blur (e.g., in the Dog and Doll sequences), the significance-guided approach maintained a stable bounding box on the target.

Qualitative tracking results on various challenging sequences. The red bounding box represents the proposed method.

Conclusion

By incorporating instance significance into the MIL framework, this research provides a more robust solution to the label ambiguity problem in visual tracking. The Bayesian ensemble approach allows the tracker to focus on the most representative samples, significantly reducing drift. Future work will focus on addressing scale variation, as the current implementation utilizes a fixed bounding box size.

Summary updated

BibTeX
@inproceedings{zhou2016online,
  author = {Tianfei Zhou and Yao Lu},
  title = {{Online Visual Tracking using Multiple Instance Learning with Instance Significance Estimation}},
  booktitle = {ICIP},
  year = {2016},
  doi = {10.1109/icip.2016.7532647},
}
Figure