Abstract
One of the main problems in face detection is the large variety of different faces. To handle faces with larger pose we introduce a new approach for pose-invariant face detection.
The idea of this project is to divide the feature space into sub spaces for reducing the complexity for hypotheses. A serialization of one detector per sub space is used. These get trained onto different pose regions and can run in sequence.
As detector we use the WVM (Wavelet Approximated Vector Machine) for its efficiency. Has the WVM a high probability, the object was found.
To optimize the performance the detectors will be arranged in a hierarchical tree structure. A root node covers the union of the sub spaces. If the root node does not reject the patch as non-object, the leaves will be traversed. The leaves are trained for smaller pose regions. The decisions of the leaves will be combined by fusion of the responses. With this approach a pose-invariant face detection can be obtained.
Key Concepts
Hierarchical Tree Structure
Detectors are arranged in a tree structure where a root node covers the union of all sub spaces, optimizing performance through hierarchical processing.
Pose Region Coverage
Leaves are trained for specific pose regions (e.g., -90° to -40°, -40° to 40°, 40° to 90°), enabling comprehensive face detection across different angles.
Wavelet Approximated Vector Machine
The WVM detector provides efficient classification. When it shows high probability, the face object is considered found in that region.
Response Fusion
Final decisions are made by combining responses from multiple leaf nodes through fusion, enabling robust pose-invariant detection.
Approach Overview
Key Result
By dividing the feature space and using a hierarchical tree of WVM detectors with response fusion, this approach achieves robust pose-invariant face detection across a wide range of face orientations.
Related Topics
Machine Learning Introduction
Learn the fundamentals of machine learning - hypothesis space, risk minimization, and VC dimension.
Download Full Thesis
Get the complete master thesis document with detailed methodology, experiments, and results.