Tree-Structured Detector for Multi-View Face Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current face detection methods face challenges in efficiently handling multi-view face detection due to variations in face poses, leading to trade-offs between accuracy and speed, and fail to effectively handle the concomitance of rotation in-plane and out-of-plane poses without increasing computation complexity or causing false alarms.
Innovation Solution
A tree-structured detector is implemented, using a width-first-search strategy and Vector Boosting algorithm to classify images, dividing face space into subspaces and employing piece-wise weak classifiers to balance accuracy and speed, while handling diverse poses through a coarse-to-fine strategy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple cascade classifiers are trained respectively for each view to handle multi-view face detection, then detection accuracy is improved, but device complexity and computation time increase significantly
Solution Approach 1:
The patent segments the face detection problem by dividing the search space into multiple subspaces, each corresponding to a specific view or pose range. Instead of training separate cascade classifiers for each view, a single cascade classifier is used with the search space segmented into subspaces that are dynamically selected during detection based on the input image characteristics.
Solution Approach 2:
The patent creates a universal cascade classifier that can handle multiple views and poses through a single detector. The classifier is designed to be multi-functional by incorporating view classification capabilities and adaptive subspace selection, eliminating the need for multiple specialized classifiers while maintaining detection accuracy across different face orientations.
2Productivity
If a pyramid structure is used to handle pose variations with coarse-to-fine strategy, then detection speed is improved, but measurement precision deteriorates due to treating different poses as one ensemble positive class
Solution Approach 1:
The patent introduces dynamic adaptivity into the detection process by enabling the system to dynamically select which subspace to search based on the input image characteristics. This dynamic selection allows the detector to adapt to different face poses and views in real-time, maintaining both high speed through selective searching and high accuracy through pose-appropriate subspace selection.
Solution Approach 2:
The patent changes the parameter of search space configuration by dividing it into multiple subspaces with different parameters (view-specific parameters). Instead of using a uniform coarse-to-fine pyramid structure, the system adjusts the search parameters dynamically by selecting subspaces matched to the detected face pose, thereby maintaining accuracy while improving speed through targeted searching.
3Productivity
If decision tree structure is used for pose estimation, then processing speed is improved, but detection accuracy deteriorates due to imperative judgments reducing time spent on pose estimation
Solution Approach 1:
The patent merges the pose estimation function with the main face detection cascade classifier. Instead of using a separate decision tree structure for pose estimation, the pose estimation capabilities are integrated directly into the cascade classifier stages, allowing simultaneous pose estimation and face detection without sacrificing accuracy through imperative judgments.
Solution Approach 2:
The patent introduces an intermediary mechanism in the form of view classification layers that mediate between the input image and the main detection process. These intermediary layers provide accurate pose estimation by classifying the face view before proceeding to detailed detection, thereby maintaining both speed and accuracy through a bridging structure rather than direct imperative judgments.
4Adaptability or versatility
If rotation invariant face detection is implemented to handle diverse poses, then adaptability is improved, but computation complexity increases significantly
Solution Approach 1:
The patent segments the rotation invariant detection problem by dividing the 360-degree pose space into multiple subspaces, each covering a specific view range. This segmentation allows the system to handle diverse poses adaptively by selecting only the relevant subspaces for each detection task, rather than processing all possible rotations, thereby reducing computation complexity while maintaining versatility.
Solution Approach 2:
The patent creates a universal detection framework that handles rotation invariant detection through a single cascade classifier with integrated view classification. This multi-functional approach eliminates the need for multiple specialized detectors for different pose ranges, achieving rotation invariance and diverse pose handling through one adaptive system rather than multiple computation-heavy specialized detectors.
Data Source
AI summary
An apparatus and a method for detecting from an image a particular subject corresponding to multiple views of the subject by dividing a particular subject space into a plurality of subject subspaces and further dividing a subject subspace into subject subspaces representing multiple views; configuring a tree-structured detector wherein the tree structure has a root node that covers all subject subspaces and has a plurality of branches, each branch corresponding to a child node that covers at least one subject subspace; training each node to determine which nodes in the adjacent lower layer the images of the subject in the corresponding nodes should be sent.