Semi-Qualitative Probabilistic Networks in Computer Vision Problems

Cassio P. de Campos, Lei Zhang, Yan Tong, Qiang JiJournal of Statistical Theory and Practice 2009 · DOI

TL;DR

This paper introduces Semi-Qualitative Probabilistic Networks (SQPNs) that combine qualitative influences with imprecise probability intervals, offering a more flexible approach than traditional Bayesian networks for computer vision tasks. They present novel learning and inference algorithms for SQPNs and demonstrate their effectiveness in facial expression recognition and image segmentation, showing improved performance in cases of interval dominance.

What Problem Does This Paper Address?

Traditional Bayesian Networks (BNs) rely on precise probability distributions, which can be difficult to elicit from experts or accurately estimate from limited data, especially in complex domains like computer vision. Qualitative Probabilistic Networks (QPNs) relax this by using qualitative relationships but often lack the quantitative precision needed for robust inference. This paper addresses the gap by proposing Semi-Qualitative Probabilistic Networks (SQPNs), which bridge this divide by incorporating both qualitative influences (like "positive influence") and imprecise probability measures represented as intervals. This allows for more expressive modeling of uncertainty and expert knowledge, particularly when dealing with situations where exact probabilities are unknown or hard to define.

Methods

The core contribution is the definition and application of SQPNs. Key methods include:

  • SQPN Definition: An SQPN is defined as a directed acyclic graph (DAG) coupled with a set of qualitative influences (e.g., positive/negative influence between variables) and range constraints (imprecise probability intervals) on conditional probabilities. This allows for a richer representation of knowledge than standard BNs or QPNs.
  • Imprecise Dirichlet Model (IDM) for Learning: To learn parameters from data while respecting expert knowledge encoded in the SQPN, the authors adapt the Imprecise Dirichlet Model. This approach combines prior beliefs (from the SQPN) with observed data to estimate posterior probability distributions as intervals.
  • Iterative Local Search (ILS) for Inference: Exact inference in SQPNs is computationally challenging. The paper proposes an Iterative Local Search (ILS) algorithm that converges to a local optimum for belief updating, maximum a posteriori (MAP), and most probable explanation (MPE) queries. This algorithm iteratively refines probability distributions for individual nodes within the SQPN constraints.
  • MAP-Reformulation: For MAP and MPE queries, a reformulation technique is introduced that transforms the problem into a belief updating query on a modified network, making it solvable by the ILS algorithm.

Datasets

  • Facial Expression Recognition: Real data from the DFAT-504 database was used. This dataset contains over 8000 frames manually labeled with Action Units (AUs) and corresponding computer vision measurements for 14 AUs.
  • Image Segmentation: Real image data was used for segmentation tasks. Specific datasets are not explicitly named, but the paper mentions using training images and their manual labeling to learn Gaussian distributions for edge measurements.

Key Insights and Ideas

  • Bridging the Precision Gap: SQPNs offer a powerful way to model uncertainty in computer vision by combining the expressiveness of qualitative relationships with the quantitative power of probability intervals, which is more practical than relying solely on precise probabilities or purely qualitative descriptions.
  • Expert Knowledge Integration: The framework effectively integrates expert knowledge (qualitative influences) with data-driven learning (IDM), allowing for more robust models, especially in data-scarce scenarios.
  • Interval Dominance as a Confidence Measure: The concept of "interval dominance" in SQPN inference provides a useful indicator of confidence. When the probability intervals for presence and absence of a state do not overlap, it signifies a strong decision, which the paper shows leads to better recognition performance compared to standard Bayesian networks.
  • Flexibility in Modeling: The ability to handle both monotonic and non-monotonic influences, and the potential to extend to other qualitative constraints like synergies, makes SQPNs a versatile tool for complex modeling.

Key Results

  • Facial Expression Recognition: In cases of interval dominance, the SQPN approach achieved higher recognition rates (positive and negative rates) compared to a standard Bayesian network trained with maximum likelihood estimation. For instance, with 1000 training frames, SQPN achieved a positive rate of 69.1% and a negative rate of 99.8% in dominant cases, outperforming the BN.
  • Image Segmentation: Visual results for image segmentation using SQPNs showed smoother contours and fewer artifacts compared to standard Bayesian networks, particularly evident in images of fish, cows, and mushrooms. The paper suggests that SQPNs can produce more visually appealing and potentially more accurate segmentations.

Limitations and Future Work

  • Local Optima: The ILS inference algorithm converges to a local optimum, not necessarily the global optimum. The paper acknowledges that the choice of initial feasible guess can influence the final result.
  • Interval Dominance Dependence: The benefits of SQPNs are most pronounced in cases of "interval dominance." The paper notes that further investigation is needed for cases where interval dominance does not hold and how to best make decisions in uncertain scenarios.
  • Future Work: The authors suggest extending SQPNs to include other qualitative constraints like synergies, exploring more sophisticated learning and inference methods, and applying them to a wider range of computer vision problems such as object and human body tracking. Further research into handling cases without interval dominance is also highlighted.

Summary updated

BibTeX
@article{campos2009semi,
  author = {Cassio P. de Campos and Lei Zhang and Yan Tong and Qiang Ji},
  title = {{Semi-Qualitative Probabilistic Networks in Computer Vision Problems}},
  journal = {Journal of Statistical Theory and Practice},
  year = {2009},
  doi = {10.1080/15598608.2009.10411920},
}
Figure