Mobile Image Recognition Pose Space Voting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition systems face challenges in identifying objects in images due to high memory requirements and processing inefficiencies, particularly on mobile platforms, where large collections of images lead to slow performance and increased memory caching issues.
Innovation Solution
The method involves offline image recognition on mobile devices using reduced pose space voting with one-dimensional accumulators, re-ranking based on spatial verification, and optimized memory access to reduce computational costs and improve speed, while also leveraging integer calculations and cache-friendly data structures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If two-dimensional accumulators are used for voting in rotation and scale space, then object recognition accuracy is improved, but memory requirements increase significantly
Solution Approach 1:
The patent extracts and removes the scaling dimension from the vote accumulator, keeping only the rotation dimension. This is achieved by performing voting only in rotation space rather than in both rotation and scale space, thereby reducing memory requirements while maintaining sufficient recognition accuracy for the application.
Solution Approach 2:
The patent applies asymmetric treatment to different pose parameters by voting only in rotation space while ignoring scale variations. This asymmetric approach recognizes that rotation requires precise voting while scale can be handled differently, thus reducing the accumulator to one dimension (rotation) rather than two dimensions (rotation and scale).
2Reliability
If large collections of reference images are processed, then recognition completeness is improved, but processing speed deteriorates due to cache flushing
Solution Approach 1:
The patent segments the large collection of reference images into smaller batches or chunks that can be processed sequentially. By dividing the large dataset into manageable segments, the vote accumulators remain small enough to fit in CPU cache during each batch processing, avoiding cache flushing while still achieving complete recognition across all images through multiple passes.
Solution Approach 2:
The patent employs periodic processing of reference image batches, where each batch is processed independently with its own vote accumulation cycle. This periodic action allows the system to clear and reset accumulators between batches, maintaining small memory footprints that fit in cache, while achieving complete coverage by repeating the process across all batches.
Data Source
Figure 1
Figure 2a~2b
Figure 3
AI summary
Methods and systems of identification of objects in query images are disclosed. Keypoints in the query images are identified corresponding to objects to be identified. Visual words are identified in a dictionary of visual words for the identified keypoints. A set of hits is identified corresponding to reference images comprising the identified keypoints. Reference images corresponding to the identified set of hits are ranked using clustering of matches in a limited pose space. The limited pose space comprises a one-dimensional table corresponding to the rotation between the object to be identified with respect to the reference image. A first subset of M reference images that obtained a rank above a predetermined threshold is then selected. Offline, hybrid and combined offline and hybrid systems for performing the proposed methods are disclosed.