Rejector Classification Tree Feature Sharing for Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object detection techniques require longer classification times and have lower efficiency due to distinct feature IDs and thresholds for each node in the classification tree, leading to a large model size and reduced detection speed.
Innovation Solution
A method for learning a rejector by forming a classification tree with shared feature IDs and thresholds across nodes of the same depth, allowing for faster classification and improved object recognition efficiency by allocating these identifiers and values uniformly within the tree.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If distinct feature IDs and thresholds are used for each node in the classification tree, then classification accuracy is improved, but classification time increases and detection speed decreases
Solution Approach 1:
The patent merges the feature ID and threshold selection process across multiple nodes by acquiring features from multiple positive and negative images simultaneously. Instead of processing each node independently with distinct feature IDs, the system combines feature acquisition from mk positive images and nk negative images, then determines a single feature ID and threshold that optimizes classification across the entire tree structure, reducing redundant processing and classification time
Solution Approach 2:
The patent implements a universal feature ID and threshold determination mechanism that serves multiple nodes in the classification tree. By using histograms of features from multiple images to decide a single feature ID and threshold, the system creates a multi-functional classification approach where one set of parameters serves multiple classification decisions, improving detection speed while maintaining accuracy
2Measurement precision
If distinct feature IDs and thresholds are allocated for each node in the classification tree, then classification precision is improved, but model size increases and processing efficiency decreases
Solution Approach 1:
The patent merges the feature representation across nodes by constructing matrices from multiple images (mk*p matrix for positive images and nk*p matrix for negative images) and using these combined matrices to determine a single feature ID and threshold. This merging approach reduces the total number of parameters stored in the model while maintaining classification precision through collective feature analysis
Solution Approach 2:
The patent creates a universal feature ID and threshold determination mechanism that serves multiple nodes simultaneously. By using histograms derived from combined feature matrices to decide a single feature ID and threshold, the system reduces model size by eliminating redundant parameters while maintaining classification precision through a unified feature representation that works across multiple classification decisions
3Measurement precision
If conventional classification methods are used with distinct parameters for each node, then detection accuracy is maintained, but processing speed decreases
Solution Approach 1:
The patent performs preliminary action by acquiring features from multiple positive and negative images simultaneously before node classification. By constructing feature matrices and determining histograms in advance, the system prepares a unified feature ID and threshold that can be applied across multiple nodes, eliminating the need for sequential feature processing and significantly improving processing speed while maintaining detection accuracy
Solution Approach 2:
The patent implements continuous useful action by using the determined feature ID and threshold across multiple nodes in the classification tree without interruption. Instead of restarting feature processing at each node, the system maintains continuous classification operation using the pre-determined parameters, improving processing throughput and speed while preserving detection accuracy through consistent feature evaluation
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for learning a rejector is provided. The method includes steps of: acquiring features from positive images and negative images of the classification tree; and deciding a specific feature ID and a specific threshold by which a minimum classification error is derived by referring to a histogram of features of positive images and that of features of negative images if a number of images classified to a k-th node is larger than that classified to a brother node of the k-th node as a result of classifying images acquired with respect to a parent node of the k-th node based on a determined feature ID and a determined threshold for the parent node and then allocating the specific feature ID and the specific threshold in nodes, which have depth equal to the k-th node of the classification tree.