Object Detector Trained Using Dynamic Working Set
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automated object detection and recognition systems face high computational costs and memory demands when training detectors on large datasets, particularly in face detection, due to the need for extensive training data to account for variations in illumination and expressions.
Innovation Solution
The system trains an object detector using a dynamic working set of training data, which is updated based on the false positive sample rate, allowing for efficient training and classification without overwhelming computational resources, and employs a Bayesian Stump to reduce the number of features required.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a large set of training data is used to train weak classifiers to recognize variations in object features, then detection accuracy and robustness are improved, but computational costs and memory demands become unacceptably high
Solution Approach 1:
The patent segments the large training dataset into multiple smaller working sets that can be processed independently and in parallel. Each working set contains a subset of training samples that are sufficient to train the weak classifiers effectively. This segmentation allows the training process to be distributed across multiple computational units, reducing the computational burden on any single unit while maintaining detection accuracy.
Solution Approach 2:
The patent applies partial action by using only a portion of the available training data (a working set) rather than the entire large dataset. The working set is carefully selected to contain representative samples that capture the essential variations in object features. This partial approach achieves sufficient detection accuracy without the prohibitive computational costs of using all available training data.
2Adaptability or versatility
If a large set of training data is used to train weak classifiers, then the detector can handle variations in illumination and expressions, but memory demands become unacceptably high
Solution Approach 1:
The training dataset is segmented into multiple working sets that can be loaded into memory simultaneously. Each working set contains a manageable subset of training samples that fit within available memory constraints. The segmentation strategy ensures that all essential variations in illumination, expressions, and other object features are represented across the different working sets, allowing the detector to maintain high adaptability while respecting memory limitations.
Solution Approach 2:
The patent uses a partial approach by loading only a working set (a portion of the full training data) into memory at any given time during training. The working set is constructed to include diverse samples that represent the range of variations in object features. This partial loading strategy reduces memory demand to acceptable levels while still providing sufficient training data for the weak classifiers to learn robust feature representations.
3Reliability
If weeks are spent training a detector with 4297 features on a training set of 4916 faces, then detection robustness is achieved, but training time becomes prohibitively long
Solution Approach 1:
The patent segments the training process into multiple stages, each working on a different working set. Instead of training on the entire large dataset in a single prolonged process, the training is divided into smaller tasks that can be executed more quickly. The segmentation allows for parallel processing of different working sets, significantly reducing the overall training time while maintaining detection robustness through the cumulative effect of training on multiple diverse subsets.
Solution Approach 2:
The patent applies partial action by training the weak classifiers on a working set (a subset of the full training data) rather than the complete large dataset. The working set is carefully constructed to contain representative samples that capture the essential variations in object features. This partial training approach achieves sufficient detection robustness in a fraction of the time required to train on the entire dataset, making the training process practical and efficient.
Data Source
AI summary
An object detector that includes a number of weak classifiers can be trained using a subset (a “working set”) of training data instead of all of the training data. The working set can be updated so that, for example, it remains representative of the training data. A decision to update the working set may be made based on the false positive sample rate—if that rate falls below a threshold value, an update of the working set can be triggered.