Object Recognition Apparatus with Accuracy-Based Vote Weighting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object recognition methods using monocular cameras face challenges in accurately estimating the 3D position and orientation of target objects, particularly when feature amounts change with viewpoint changes, leading to difficulties in deciding vote weights for reliable recognition.
Innovation Solution
A learning method that selects specific regions from a 3D model, learns detectors for these regions, evaluates recognition processing, and normalizes vote weights based on recognition accuracies to prioritize votes from highly reliable region sets, thereby improving recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If vote weights are decided discretely from pre-prepared weight variations, then the search speed is improved, but the granularity of weight variations deteriorates
Solution Approach 1:
The patent changes the parameter representation from discrete pre-prepared weight variations to continuous vote weights calculated based on recognition accuracy. This allows for finer granularity in weight adjustments while maintaining computational efficiency through the accuracy-based calculation method.
2Adaptability or versatility
If feature amounts are used for objects at arbitrary orientations, then the adaptability is improved, but the reliability of vote weights deteriorates due to viewpoint-dependent feature changes
Solution Approach 1:
The patent introduces feedback by calculating recognition accuracy for each region set and using this accuracy as feedback to determine the vote weight. This closed-loop approach ensures that vote weights reliably reflect the actual performance of each region set, even when feature amounts vary with viewpoint.
Solution Approach 2:
The patent changes the parameter used for weight determination from static pre-prepared weights to dynamic weights based on calculated recognition accuracy. This adaptation allows the system to maintain reliable vote weights across arbitrary orientations by basing weights on actual performance rather than assuming fixed characteristics.
3Device complexity
If all region sets are treated equally in voting, then the simplicity is maintained, but the recognition accuracy deteriorates due to unreliable votes from low-quality region sets
Solution Approach 1:
The patent applies local quality by assigning different vote weights to different region sets based on their individual recognition accuracies. Instead of treating all region sets uniformly, each region set receives a weight proportional to its reliability, thereby improving overall recognition accuracy while maintaining a relatively simple voting framework.
Data Source
AI summary
A learning method of detectors used to detect a target object, comprises: a selection step of selecting a plurality of specific regions from a given three-dimensional model of the target object; a learning step of learning detectors used to detect the specific regions selected in the selection step; an evaluation step of executing recognition processing of positions and orientations of predetermined regions of the plurality of specific regions by the detectors learned in the learning step; and a normalization step of setting vote weights for outputs of the detectors according to recognition accuracies of results of the recognition processing in the evaluation step.