Parts-Level Detector Training with Verified Image Timelines
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for object detection require labor-intensive and resource-heavy dataset generation, with dataset quality significantly impacting prediction accuracy, necessitating improved input datasets and training methods.
Innovation Solution
A method and system for generating a detector model that involves creating verified training datasets by discarding images without corresponding ground truth tags, using a two-level detector training process with parts-level detectors and a unified detector, reducing the need for exhaustive tagging and resource burden.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a comprehensive training dataset is created with exhaustive object tagging, then detection accuracy is improved, but the time and labor required for dataset creation increases significantly
Solution Approach 1:
The patent segments the training dataset into multiple subsets, each focused on specific object classes or detection scenarios. Instead of creating one comprehensive dataset with exhaustive tagging, the system divides the data into targeted subsets that can be processed and validated more efficiently, reducing overall dataset creation time while maintaining detection accuracy for specific object classes
Solution Approach 2:
The patent applies partial action by creating training datasets that are sufficient rather than exhaustive. The system identifies and tags only the critical object classes needed for specific detection tasks, rather than attempting to tag all possible objects in the scene. This partial tagging approach reduces the time and labor required for dataset creation while maintaining adequate detection accuracy for the target objects
2Reliability
If manual object tagging is performed on all images, then dataset quality is improved, but the labor intensity and resource burden increase
Solution Approach 1:
The patent implements self-service by using automatically generated ground truth timelines from detector outputs to verify and validate training datasets. The system uses the detector's own predictions to create initial ground truth data, which then serves as a basis for automated verification. This self-referential approach reduces manual labeling effort while maintaining dataset quality through automated consistency checks
Solution Approach 2:
The patent incorporates feedback mechanisms where detector outputs are used to generate ground truth timelines that are then fed back into the training process. The system continuously refines the training dataset by comparing detector predictions against ground truth timelines and automatically correcting inconsistencies. This feedback loop improves dataset quality without requiring exhaustive manual tagging, as the system self-corrects through iterative refinement
3Measurement precision
If ground truth timelines are generated manually for verification, then false positives and negatives are reduced, but the processing time increases
Solution Approach 1:
The patent replaces manual mechanical processes with automated computational methods. Instead of manually creating ground truth timelines for verification, the system uses automated algorithms to generate timelines from detector outputs and systematically compare them against training data. This substitution of manual mechanical tagging with automated computational verification maintains detection precision while dramatically increasing dataset processing speed
Solution Approach 2:
The patent applies preliminary action by generating ground truth timelines in advance using automated processes before the verification step. The system pre-processes the training data to create initial ground truth references that can be quickly compared against detector outputs. This preliminary automated generation of verification data eliminates the need for time-consuming manual timeline creation during the verification phase, maintaining precision while improving processing speed
Data Source
AI summary
A method of generating a detector includes obtaining a first training dataset having a first image sequence with a first set of object tags identifying at least one first object class in a corresponding image. A first set of ground truth tags is obtained based on a ground truth timeline identifying when the at least one first object class appeared in the first image sequence. Images from the first training dataset are discarded by either identifying object tags by class from the first set of object tags without a corresponding ground truth tag from the first set of ground truth tags or identifying object tags by class from the first set of ground truth tags without a corresponding object tag from the first set of object tags to generate a first verified training dataset. A first parts-level detector is trained based on the first verified training dataset.