Object Recognition Using Exemplar Descriptors and Geometric Voting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object recognition techniques in images are inefficient and computationally expensive, especially when dealing with small intra-class variations and varying lighting conditions, and they struggle to detect objects from different directions.
Innovation Solution
A process that uses n-dimensional vectors to represent pixel properties in subregions of images, with exemplar descriptors selected through interest point detectors or dense sampling, and employs re-lookup tables to efficiently calculate visual similarity coefficients, allowing for precise object recognition under various transformations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If parametric classifiers are used for object recognition, then classification performance is improved, but computational cost and time complexity increase with the number of classes
Solution Approach 1:
The patent segments the recognition process into two stages: first, a fast coarse classification stage using histogram comparison that quickly eliminates non-matching classes; second, a fine-grained verification stage using geometric layout verification only for candidate classes. This segmentation reduces the computational burden of verifying all classes while maintaining high classification accuracy.
Solution Approach 2:
The patent performs preliminary histogram computation and comparison before geometric verification. By pre-computing histograms and performing fast histogram-based filtering, the system prepares candidate classes in advance, so that computationally expensive geometric verification is only applied to a reduced set of promising candidates, improving overall efficiency.
2Measurement precision
If geometric layout verification is performed, then object recognition accuracy is improved, but computational time and complexity increase
Solution Approach 1:
The patent divides the verification process into hierarchical levels: histogram-based coarse verification followed by geometric layout verification only for candidate matches. This segmentation ensures that geometric verification, which is computationally intensive, is applied selectively rather than to all possible classes, reducing time loss while maintaining accuracy.
Solution Approach 2:
The patent performs preliminary histogram comparison to identify candidate classes before applying geometric verification. This preliminary filtering step reduces the number of geometric verifications needed, thereby reducing computational time while preserving recognition accuracy through the subsequent precise geometric checking of candidates.
3Stability of the object's composition
If aggregated representations are used for image classification, then class imbalance is reduced, but image-to-image matching precision is lost
Solution Approach 1:
The patent segments the matching process into two parts: aggregated histogram comparison for robust class-level classification that handles class imbalance, and detailed geometric verification for precise instance-level matching. This segmentation allows the system to benefit from both the stability of aggregated representations and the precision of detailed comparison.
Solution Approach 2:
The patent uses aggregated histogram representations for preliminary class classification, which provides robustness to class imbalance. Then, for identified candidate classes, it performs precise geometric verification to recover the matching precision that would otherwise be lost in aggregation, achieving both balance and precision.
Data Source
AI summary
A computer implemented recognition process of an object in a query image provides a set of training images, each training image being defined by a plurality of pixels and comprising an object tag; determines for each training image of the set a plurality of first descriptors, each first descriptor being a vector that represents pixel properties in a corresponding subregion of the associated training image; and selects among the first descriptors a group of exemplar descriptors describing the set of training images, wherein selecting the exemplar descriptors includes determining the first descriptors having a number of repetitions in the set of training images higher than a certain value.