Feature Integration with Adaptive Importance Maps for Visual Tracking

Aishi Li, Ming Yang, Wanqi YangIJCAI 2018 · DOI

Introduction

Visual tracking is a fundamental computer vision task that requires balancing accuracy with real-time performance. Discriminative Correlation Filters (DCF) have emerged as a dominant paradigm due to their efficiency in the Fourier domain. To enhance robustness, modern trackers often integrate multiple features, such as Histogram of Oriented Gradients (HOG) for shape and Color Names for appearance. However, the standard approach of simply concatenating these features into a single multi-channel block often leads to overfitting and fails to exploit their complementary strengths. This paper introduces Correlation Filter with Weighted Feature Integration (CFWFI), a framework that dynamically weighs features through adaptive importance maps.

Methodology: Adaptive Importance Maps

The core innovation of CFWFI is the joint learning of tracking filters ($h$) and a set of Adaptive Importance Maps ($m$). Unlike static weighting schemes, these maps are updated frame-by-frame to emphasize the most discriminative spatial regions for each specific feature type.

Figure 4: Visualization of the learned importance maps (a) and the resulting response maps (b) for different feature channels. The importance maps act as a spatial attention mechanism, highlighting reliable regions for HOG and color features respectively.

Key technical components include:

  • Channel Sharing: To prevent overfitting, all channels within a single feature (e.g., the 31 channels of HOG) share a single importance map. This forces the model to learn general spatial reliability rather than noise in individual channels.
  • Regularized Optimization: The authors introduce an $L_2$ regularization term for the importance maps, controlled by a penalty factor ($\alpha$). This allows the tracker to prioritize generally more reliable features like HOG over intensity maps.
  • ADMM Solver: To maintain speed, the complex objective function is solved using the Alternating Direction Method of Multipliers (ADMM), which decomposes the problem into efficient sub-problems solvable in the frequency domain.

Experimental Results

The CFWFI tracker was evaluated on the OTB13 and OTB15 benchmarks, which include challenging scenarios such as occlusion, fast motion, and illumination variation.

Figure 1: Qualitative tracking results across various sequences from the OTB benchmark. CFWFI demonstrates superior robustness in scenarios involving target deformation and lighting shifts compared to baseline DCF methods.

Performance Metrics

  • Accuracy: On OTB13, CFWFI achieved an Area Under Curve (AUC) score of 0.695, significantly outperforming baseline trackers like BACF (0.648) and Staple (0.593).
  • Real-Time Speed: Utilizing handcrafted features (HOG + Color), the tracker maintains a speed of 22 FPS on a standard CPU, making it viable for practical applications.
  • Deep Integration: A variant utilizing CNN features (DeepCFWFI) achieved a precision score of 0.854 on OTB15, though the computational cost reduced the frame rate to 9 FPS.

Conclusion and Future Work

CFWFI demonstrates that intelligent feature integration is as critical as the features themselves. By visualizing the importance maps, the authors show that HOG maps focus on high-frequency edges while intensity maps prioritize homogeneous regions, allowing the tracker to adapt to environmental changes like illumination shifts or deformation.

Future research directions include optimizing the integration of deep features to reach real-time speeds and automating the selection of penalty factors through meta-learning. Additionally, incorporating a long-term re-detection mechanism could further improve robustness against total target loss during extended occlusions.

Summary updated

BibTeX
@inproceedings{li2018feature,
  author = {Aishi Li and Ming Yang and Wanqi Yang},
  title = {{Feature Integration with Adaptive Importance Maps for Visual Tracking}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/108},
}
Figure