Hard Negative Proposer for Object Detection Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object detection and recognition technologies face inefficiencies in learning due to repeated bootstrapping processes with numerous negative images, leading to increased operational time and the need to selectively learn only hard negative images mistaken for positives.
Innovation Solution
A method involving a hard negative proposer that identifies and transmits hard negative image candidates to a classifier, using a similarity map for semantic and visual similarity, to selectively acquire and classify additional hard negative images, thereby optimizing the learning process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a lot of negative images are used for bootstrapping process, then the classifier can learn more comprehensive negative cases, but the operation quantity increases too much resulting in too much time for whole learning
Solution Approach 1:
The patent extracts only the most valuable subset of negative images (hard negative images) from the entire negative image set. By identifying and selecting only those negative images that are most similar to positive images and most likely to be misclassified, the system removes unnecessary negative images from the learning process, thereby reducing learning time while maintaining classification accuracy
Solution Approach 2:
The patent changes the selection criterion from using all negative images to using only hard negative images identified through similarity measurement. By introducing similarity metrics and selecting negative images based on their proximity to positive images in feature space, the system transforms the learning process to focus on critical cases, reducing overall learning time while preserving accuracy
2Productivity
If only hard negative images mistaken for positive images are selected for learning, then the learning efficiency improves, but it becomes difficult to identify and select the appropriate hard negative images from the large number of negative images
Solution Approach 1:
The patent implements a feedback mechanism where the classifier's predictions are used to identify hard negative images. The classifier processes negative images and provides feedback on which ones are most likely to be misclassified as positive. This feedback loop enables automatic identification of hard negative images without manual intervention, improving learning efficiency while solving the identification difficulty
Solution Approach 2:
The patent introduces similarity measurement as an intermediary mechanism between negative images and positive images. By calculating similarity metrics between negative images and positive image features, the system automatically identifies which negative images are most likely to be misclassified. This intermediary measurement process solves the difficulty of identifying hard negative images from large sets
Data Source
AI summary
A method for supporting to collect hard negative image is provided. The method includes the steps of: (a) a classifier receiving from a hard negative proposer, and classifying, hard negative image candidate corresponding to a certain label candidate or a specific label candidate judged to have semantic or visual relationship with a target object; and (b) the classifier (i) allowing the hard negative proposer to select an additional label candidate whose similarity to the certain label candidate or the specific label candidate exceeds a preset degree of similarity, if a percentage or a number of the hard negative image candidate mistaken for having the target object is judged to satisfy a prefixed condition, (ii) receiving at least one additional hard negative image candidate corresponding to the selected additional label candidate and (iii) classifying the additional hard negative image candidate.