Bayesian Network Based Reparameterization of Haar-like Feature

Hirotaka NiitsumaAAAI 2006

Abstract

The classic Viola-Jones object detection framework, while revolutionary for its speed, remains limited by its rigidity regarding object scale and orientation. This paper proposes a formal probabilistic framework that reparameterizes Haar-like features within a Bayesian Network (BN). By modeling the spatial relationships between features through affine transformations, the detection process is reformulated as a probabilistic inference task. This approach allows for the simultaneous estimation of an object's position, scale, and rotation without the need for exhaustive retraining or multi-scale scanning windows.

Introduction

Since its introduction in 2001, the Viola-Jones detector has been a cornerstone of real-time object detection, particularly for faces. Its success stems from the use of Haar-like features—simple rectangular filters—processed through a boosted cascade. However, these features are traditionally axis-aligned and static. Detecting rotated or scaled objects typically requires either training multiple detectors or employing a computationally expensive sliding window approach across various scales and orientations.

This research addresses the lack of a formal probabilistic model for these features. The goal is to describe an object as a structured collection of features that maintain their relative geometric relationships under affine transformations. By doing so, object detection evolves from a binary classification problem into a parameter estimation problem.

Methodology

Conditional Haar-like Features (CHIF)

The first innovation is the transition from static pixel-sum differences to Conditional Haar-like Features (CHIF). Instead of a fixed filter, a feature is approximated using three vertices and relative location parameters. This is formulated as a conditional probability:

$$P(z|\hat{x}) = \mathcal{N}(\mu + \Lambda\hat{x}, \Psi)$$

where $z$ is the feature response and $\hat{x}$ represents the local geometry. This reparameterization allows the feature's response to be predicted based on its relative position within an object's frame of reference.

Reparameterization of a Haar-like feature using vertices and relative location parameters.

Bayesian Network Representation and Variable Translation

An object is modeled as a Bayesian Network where nodes represent individual CHIFs. To facilitate detection across different coordinate systems, the authors introduce a variable translation technique. A transformation matrix $\Omega$ maps the coordinate system of the observed image back to a "registered" model coordinate system.

The relationship between the absolute location $x$ in the image and the relative location $\hat{x}$ within the detection window $u$ is governed by these transformations, allowing the model to mathematically "undo" rotation and scaling during the inference process.

Geometric relationship between absolute image coordinates, window location u, and relative feature locations.

Detection via Inference

Detection is performed by substituting the affine transformation into the BN and calculating the likelihood of the object's parameters $u$ (position, angle, and scale). The optimal parameters are found by maximizing this likelihood:

$$\hat{u} = \arg \max_u p(u | \text{observations})$$

This allows the system to "infer" the exact state of the object rather than just confirming its presence.

Experimental Results

The framework was evaluated using a controlled setup involving synthetic images subjected to various rotations, scales, and shifts. The results indicate that the system can successfully recover the target's position, scale, and rotation. Specifically, the inference remains robust as long as the sampling area's center is within approximately half of the target's size from the true center. This demonstrates a significant "capture range" for the probabilistic inference compared to rigid template matching.

Discussion and Limitations

While this approach provides an elegant probabilistic foundation for Haar-like features, it introduces new challenges. Bayesian Network inference is generally more computationally demanding than the highly optimized integral image calculations used in the original Viola-Jones framework. Furthermore, the current experiments rely on manually constructed BNs for simple geometric shapes. Scaling this to complex, high-variance objects like human faces in unconstrained environments would likely require an automated learning phase, perhaps utilizing Expectation-Maximization (EM) to learn the BN structure and parameters from large datasets.

Summary updated

BibTeX
@inproceedings{niitsuma2006bayesian,
  author = {Hirotaka Niitsuma},
  title = {{Bayesian Network Based Reparameterization of Haar-like Feature}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}

Loading related content...

Figure